Is it that claudish is seeping into the common vernacular, or that I'm becoming hypersensitized to certain turns of phrase? There are a handful of paragraphs that read like LLM text.
No this screams AI written to me as well. Not so bad as to have not been edited, but the sentence structure reads very heavily as AI. Likely with significant editing.
Analysis by Pangram suggests that this text was written with some AI assistance."
That's what comes up for me. Once logged in to Substack (on PC), I had to go to my account -> subscriptions -> find the post there -> click on it -> then the three dots showed up and I could scan for AI.
I don't know that I necessarily trust these types of "scan for AI" results, but that's what it's saying.
> Note: The AI transparency tool is unavailable for video or audio posts, or for posts viewed on standalone Substack sites (such as on.substack.com or custom domains) or emails.
Weirdly, I’ve only been able to see it on my own articles, not on others! It’s only visible for me in the format where a Substack article seems to sit in a window in front of a bigger Substack background window, when comments appear in truncated form rather than with full threading.
I remember having seen this on the site, but I didn't write it down on my list. I was confused at first, because the whole concept seemed promising, and then the LLM speak started. Such a shame.
From "without naming where I'm coming from" onwards, this had so much AI it was unreadable. I came to the comments to see if others had noticed too - thankfully yes! The majority of this reads like AI to me.
Guess this is (indirectly) my fault; I read all the finalists beforehand to make sure they were publishable, and I didn't notice anything wrong with this one. I guess I'm worse at detecting AI than most of you.
(I was possibly biased because there was a minor incident involving a breach of anonymity with this one, and it was by a person who I thought didn't usually use AI. Also, I'm a sucker for any review that mentions the Amateur Transplants)
It didn't read like AI to me, so count me in as also one of the incapable at detection. Art is a lot easier to identify, and those interminable slop articles which go on forever, have you endlessly scrolling, and repeat paragraphs over and over without coming to a conclusion, but this post read enough like human-written that nothing leaped out at me as "okay, AI phrasing here".
"It's not this, it's that" seems like a very hard pipe for these things to put down -- I didn't count, but there's several instances; it's "better" than it used to be I guess in that there's less of that exact phrasing, but the semantic pattern is still there a lot.
I wonder why that is? It's actually not a super-common thing to do in normal English...
I think there could be multiple explanations for the not x, but y pattern. It's a way to condense a whole argument into one sentence, which saves space. It's also a way of framing things that seems subtly intent on "blowing people's minds". I think part of it is that these things were designed (or learned?) to build their own hype, so that they seem more useful/valuable to their customers if they are capable of making new revelations or overturning people's understandings.
Likewise. But then, I've read a lot more of the kind of writing AI is trained on than I've read actual AI output. My radar isn't well-tuned at all.
That said: I've also had the experience of seeing people decide that a piece I wrote 100% myself was AI, which leaves me more than a little bit skeptical about the LLM police.
I also did not suspect this might have been LLM-written. The only passage that struck me as odd was “The patient is the doctor. The doctor was the patient. The physician could not heal himself,” where the first two sentences repeat each other.
As for authors trying too hard to be cute, _The King’s English_ (1906) covered that in the chapter “Airs and Graces.” The particular cutenesses may have changed, but the problem hasn’t.
For me, the main indications of Claude were the uses of “name” as a verb, and the discussion of what “nobody” would do. It only struck me in a few passages, but it was very distinctively never B’s passages, just A’s.
I am fairly confident that A used Claude, and in my opinion the most obvious tell is a repeated tendency to frame things in a very particular, black-and-white sort of moral language which heavily relies on phrasing and the judicious use of articles to create its argument. I would say one of the best examples is:
>The consultant is at home. The SHO is calling Kay. The midwives are calling Kay. Five things are happening at once. A decision that will be reviewed in a coroner’s court two years later is being made by a sleep-deprived doctor whose nominal supervisor is asleep across the city.
Yeah, I think A’s side may not be AI-written, but very much feels AI-edited. I think we’ll see much more of this going forward - LLMs are now starting to be used the way dumb spelling+grammar checkers used to be.
I noted that breach as well and have been curiously waiting for this review to drop -- the primary reason being that the person in question is a huge and unrepentant LLM user who was bragging about how this finalist selection meant that his LLM usage was perfectly reasonable!
I am not good at detecting AI either, and I'm pretty picky about writing. In fact I have to say that I *like* a lot of AI writing. I discuss ideas and theories with AI pretty often, currently mostly with Astra, and it frequently happens that I ramble on for a long paragraph about my thoughts on a subject, and Astra's response summarizes my main point in a clear, smart way, using phrases *better than any I came up with.* She does are often display some irritating AI mannerisms -- for instance, flattery of me, the over-frequent use of phrases like "it's important to distinguish between . . ." But that's not nearly bad enough to neutralize my pleasure in Astra's crisp, clever takes on what I say.
I'm not sure what to make of this pile-on of complaint about the Claudishness of A. Frankly, one reason I am not very enthused about ACX book reviews is that a lot of them are badly written. Reviewers ramble on about the book without any central idea they want to get across about the it. Their prose is blurry. They make grammatical errors I don't have to scan for to notice -- mistakes that bother me like sand in a salad. They use the non-word "irregardless." They don't serve up any delightfully apt ways of putting things. Claude or Astra prose seems better to me than at least 90% of ACX book review prose.
Is this sort of a fad, hating on Claudishness? Or am I just missing some way that Claude's communications are ugly and impoverished?
I'm in favor of our having a rule against AI-written posts or book reviews, but for me that's because I come here to read what *people* think, not because I believe AI communications are intrinsically low-quality. I don't think they are.
Good human writing uses high-intensity rhetorical techniques like school students use a highlighter, while AI writing uses them at inappropriate times that don't make any sense.
The classic not-this-but-that structure is ridiculous when it's applied to the things AI writes about: "Not butter. Not mustard. Just honest mayonnaise." Why do we care about the mayo in particular? We don't! Why is it honest? It isn't! And yet the AI has added a few negations and slapped down a useless adjective because that's a common pattern in good writing, where human writers use it to hammer a central point into the mind of the reader.
I find the style fatiguing to read. I don't understand how other people are able to look at it without thinking "get to the point" and finding their eyes skip over the text like it's an ad on a website. When I'm using Fable, I have to instruct it to process its own output into plain English, and even then it's an unreadable mess of "clever" obscurantism.
For me, there’s just a slight annoyance in recognizing AI composition of text that is presented as coming from a person. But yes, it’s better than a lot of human writing.
It's not exactly that Claude's communications are ugly and impoverished (though, through some lenses they are), it's that they're _so obvious_ to those of us who see them.
Additionally, while its certainly possible to produce insightful and useful writing that is largely written by AI, the vast majority of AI writing you encounter on the internet is just slop.
So, couple these together:
* some people seem to be able to identify AI writing quickly and easily
* AI repeats the same grammatical tropes _constantly_ so as to make it an irritating habit
* most AI writing most people encounter is just slop or spam or had no author involvement other than "write about X and post it".
And yeah, people hate it when they encounter AI writing.
I strongly believe that this entry should be disqualified and another nominee should be added to take its place.
As a writer, albeit one who did not enter this contest, I really don't like seeing AI writing (or half-AI writing) taking slots from real human beings.
This raises the question of whether or not AI use is disqualifying. You kind of imply you were looking for it, but it's not mentioned explicitly in the rules.
For my part I will be strategic voting against this entry as hard as strategy permits.
I noticed it pretty quickly with phrases such as “and it has been failing in ways that compound”. After that, I was expecting to read that something was "load-bearing".
Adding one more to the pile. "Both readings have things going for them. Both miss what I think the book actually does, and by a country mile." was the point I had to jump ship.
I'm 95% sure this was written by a fellow going by the name (elsewhere) of self_made_human, who openly professes to use LLMs to write the majority of his comments. You can safely assume this was written similarly.
Sounds like AI to me. I'm not categorically opposed to it in principle, but I find that in practice it comes with unnecessary repetition and summarization. The AI isn't good at killing its darlings, and a lot of the little grace notes AI seems to like don't really land ("amputation of what falls off"?). AI is just not good at figures of speech. Overall, it ends up being pretty tiring to read. Over the course of this review, I went from reading to skimming to scrolling just to get to the end. Not to say I don't do that with obviously human writing sometimes as well.
I've done a lot of work with AI for both coding and research/writing tasks, so maybe I've been particularly sensitized to its quirks. It tends to use metaphors and other figures of speech that don't really work, but are somehow related to the topic, quite often. So in that paragraph it references ischaemia and amputation while discussing the healthcare system, which to me is a clear tell. If you look over the review, I bet you'll find similar usage in a number of places.
There was obviously a lot of editing done after Claude. There’s multiple sentence constructions that are literally one Claude authored sentence and that, and then a human authored section.
I get that different people have massively different levels of sensitivity to reading LLM-authored prose. I'm one of the people in a group (whose size I do not know) that feels something close to disgust, but directionally pointed more towards frustration when reading works with the LLM-isms.
But to those people who say hey it’s no problem. It’s perfectly fine to you know have stuff where there’s LLM written sections. All I can say is fuck you and your tribe.
Agreed. I know that some people have poor writing skills and/or fluency in English, and in principle I'm interested in their ideas, even if they need an AI to express them. But in practice, the second I clock even one sentence as AI, my brain revolts against reading the whole piece.
Next year the AIs will be better, and we won't be able to easily sniff them out. But I expect I'll still have the same disgust reaction, and it will extend to all writing that isn't certified human. So I sadly do want a social norm against AI writing - otherwise I soon won't have much to read.
Not sure. I’d be right there on the “disqualify” side if it was unedited or lightly edited LLM output, but this doesn’t feel like that. It reads mostly human with a light dusting of LLM - just enough to be jarring, I guess. I’d much rather the author felt confident enough in their own writing to not replace it with LLMisms, but is that a sin worth disqualification?
I want to read the output of a human brain, and if they use software tools to help them write I think that is bearable. I do not want to read AI slop, even when the AI uses a vestigial human to mask itself. I don’t know where the line should be drawn, or even if we can draw a better bright line in a competition format where whatever rule we are able to articulate is sure to then be gamed, but as LLMs become more useful tools I am unconvinced that literally 0% LLM is the right place to draw the line.
Scott, would it be at all possible to arrange a poll like the art one from a while back, but with text? 5-10 sentence or so sections on some subject, contributed by humans, written by LLM, tool-assisted human, human-tweaked LLM; then we could all calibrate our LLM detectors and see if we actually are as good at picking out the clankers as it feels like. I don’t imagine it’d put this stuff to rest, but we would at least have some data to put behind future conversations, and if you ever wanted to put together a panel of people better than average at telling which was which this would generate some leads. It may even warrant repeating annually or biannually - I can imagine the results shifting as models improve while Claudisms enter human discourse.
Outstanding. I hadn't read Kay's book for the reasons which A and B adequately give - Kay sounds like a very difficult person/writer/doctor to like - but I appreciate their nuanced assessment, finding the vulnerable, even the caring, man among the cliches and gags. Surprised by other commenters getting AI vibes - I often do, but didn't here.
There was a good essay in here before it got fed into whatever LLM dragged it out to three times the length it needed to be. It's the rhythm of it that really grates after a while.
A lot of this review is very clearly written by AI, particularly this paragraph: 'This Is Going To Hurt is, read with a clinical eye, a textbook case study of moral injury and post-traumatic stress in a competent doctor, narrated by the patient himself, with the diagnosis hidden in plain sight. Kay is both author and index case. The book is his presenting complaint. The institution that produced him then declined to treat him, and the book is the receipt for that decline. Once you see the book this way, several things follow. Kay’s defenders are right about more than they realise; his critics are wrong about more than they think; and the system that broke him in 2010 is, in 2026, still producing more of him, faster.' which is the most extreme claudish I've ever seen.
What exactly is bad about it? Are you saying it's incoherent? I don't think it is. I'd be able to summarize the points in it in other language. Wouldn't you? I do see how it's trying a bit too hard to be clever by collapsing levels on each other: Kay the writer and Kay the index case, the book as the book and the book as natural result of the institutional failure it describes. And there's something facile and not altogether valid about some of the mental moves it's making. So I'm not crazy about that paragraph either. On the other hand, I enjoy and respect the *kind* of cleverness it's trying to engage in here, even though the version of it in the paragraph you quote is definitely second rate. A lot of people can't even do second-rate versions of this. And I've seen Claude do first-rate versions, just not here.
Yes, it’s very competently written by human standards, but the way Claude (Opus and Fable in particular) constructs sentences when it’s trying to be artful tends to result in stuff like this, where the phraseology is stacked up on itself in compounding metaphors that are introduced in parallel and all leading back to the same referent (is the book a receipt for decline or his presenting complaint? Etc). That’s what I mean by spaghetti prose. Each individual spaghetto can be traced, it’s just more work to do that than it should be when the goal of this sort of writing is to communicate an idea or theme clearly.
I say this all as someone who uses Claude all the time and finds it immensely useful, including for writing. I think familiarity can breed contempt in that respect, too. Once you encounter this particular type of cleverness for the second, third, or 500th time it starts to wear very thin.
The point of reading is to mind-meld with the author. But when it comes to Claude’s writing, you’re mind-melding with an author that constantly gets in the way of their own message because they can’t help but stop every few sentences to be *cute*.
I have read other AI-generated prose online and yes, it does try to be too clever when it is striving to sound artful. Over-wrought, too descriptive, and verbose. A lot of metaphors and similes that in the end are meaningless filler.
(I, er, may also indulge in some of this in my own writing so glass houses, stones, you know the drill).
> Kay’s defenders are right about more than they realise; his critics are wrong about more than they think
Really look at that phrase. What is it saying? “People I agree with are right, and yet somehow don’t realise it. People I disagree with are wrong, but they think they are right.” It is a nonsense bit of writing. It is “heads I win, tails you lose” dressed up as if it is deeper insight. Why would his critics think they were wrong about anything?
I don't think it's nonsense, exactly. There could be a situation where critics are saying, "Ok, maybe I'm placing too much weight on A, but B, C and D I am sure are solid evidence of my point." So there could be a situation where critics are acknowledging that a few of their complaints my be unjustified, but in fact ALL their complaints are unjustified. But I gather this is not one of those situations. In fact it's probably often the case that people who loathe something have a mental list of every criticism anyone has ever made of the thing, and trot out the full list in all their attacks, while being half aware that some of them are weak and not really defensible. And it would not be surprising if the reverse of all this is often the case for people that are in favor of something.
Still, there's no reason given for thinking there's an unusual or interesting amount of that going on about Kay, so I agree that phrasing the point that way is really just a stylistic flourish masquerading as a description of the structure of the disagreements about Kay.
"His defenders don't go far enough, because there are other things they don't know or have not considered that belong to the argument they are making. His critics are correct in some of their criticisms, but wrong because they are criticising the wrong thing or don't have the full picture."
Here is, I think, how the feeling of glossing over llm-prose goes for some people.
Sure you can construct a reason that it makes sense. But if it was substantially not written by the author why should I bother to make sense of it ? I can construct a meaning out of almost anything if I bother to, shapes in the clouds, tarot cards, the I Ching, but that's all junk and the meanings would lead me astray except if I used them as toy randomizers for my mind.
Llm's clearly produce output that is more meaningful, more true, perhaps more interesting, than divination rituals, but this doesn't quite help. Lots of people don't read to hear just the interesting or more true idea. They are interested in knowing the mind that produced the idea. If this is what you're expecting from writing than the author delegating to Claude is at worst deceit (passing off the ideas of Claude as their own) and at best a waste of that readers time (Claude is either/or both not a person and not a person worth understanding).
Anyone reading something for the purpose of understanding another person will be pissed off with you no matter how much you protest that the writing is better.
Somewhat less loftily there is also some amount of straight cheapening going on here. Everyone has to discriminate in what they're doing with their time. Claude costs effectively nothing and people sure seem to think it increases their productivity in producing texts. So then when someone recognizes that they are reading Claudisms another reaction (which is I think often justified) is "Why are you serving me porridge oats ? If I wanted to eat peasant slop for pennies a meal I would have stayed at home."
Putting aside the AI question, as I have no strong opinion, though especially towards the end, several sections felt a bit like that, frankly, I was wondering if one of the speakers was AI instructed to act as an adversary to the prior speaker? But though a lot of this is quite good and effecting, the defense against misogyny is...incredibly weak?
Basically irrelevant sections are held up as dispositive? I haven't read the underlying claim and the point about his patients being uniformly female is a reasonable one, but 'he was affected and cared about a case with a dead baby' and 'he was good to his friend's dad when he was dying of cancer'...don't have anything to do with the claim of misogyny? Like, if the claim was misanthropy, I'd understand?
Glad I wasn’t the only one who noticed this! There is zero tension between being misogynist and liking babies and old men WTF. But maybe this is the sort of thing you don’t catch if you don’t do the writing yourself…
Ah, yeah. "He was kind to someone with whom he had a personal relationship" does not make "he has a pretty nasty view of women" harder to make in good faith.
"46 For if you love those who love you, what reward do you have? Do not even the tax collectors do the same? 47 And if you greet only your brothers, what more are you doing than others? Do not even the Gentiles do the same?"
Yeah, that's what I was thinking too. Not knowing who A or B are, I thought that A might have been completely written by an LLM instructed to be an adversary for B.
Yes, it’s like they think “misogyny” means “an irredeemably evil person in every aspect of their being, motivated only by pure hatred and selfishness”. Which I suppose is a more understandable position to take if you are a LLM trained on internet arguments.
Tough read. Not just AI but very repetitive and overly verbose. Genuinely surprised this is a finalist without the author asking all their friends (“friends” possibly being their own emails) to vote for them or something.
While I think that medicine is not doing its best regarding the treatment of women's issues?illnesses?complaints?, women, on the other hand, expect magic, and complain *bitterly* when they don't get it. They don't know how their own reproductive organs work. They don't understand pregnancy and childbirth as natural processes. They don't research things that would be helpful. They don't avoid things that are unhelpful (being a slag for instance). They are unwilling to take responsibility for the decisions they make that degrade their own health and well-being. I cannot imagine being anything other than cynical if you had to work almost exclusively with female patients.
As for socialized medicine, like public education, it is clear that it does not work. Caring for either the health of the individual or the developing mind of the individual does not scale. And ever pretending it did was a snafu which eventually only benefited those who care for neither the health of the individual nor the developing mind of the individual.
I said it doesn't scale. The US is a huge country, and whatever country you are talking about could probably be dropped into the middle of the US and disappear.
And socialism only works when you have a lot of money to steal. Once you have more money going out than coming in, things get dicey.
As for women complaining, geez - oldest story in the world, literally.
Hmm, but it is currently running out of money. Demography is catching up, and the pyramid scheme is starting to fail very, very badly. It is actually about 10x worse then our German pension problem.
But like many social systems, it was built on assumption that were always true (growing population and a proper demographic pyramid) until right about now. Part of its downfall is it's own success (extending lifespan radically), but the larger problem of demography is outside of it.
It's not a critique of the system per se, just injecting some current trends into the discussion. I live in it and I know the details, if you have more questions.
Hmm, I don't see the relevance. I'm not an advocator of private insurance as a solution, if that was your interpretation. I did not mean to take part in your private-vs-public discussion with MLHVM on his side.
I don't think that any insurance relies on demographics per se. A relatively low price might rely on demographics, but in principle insurance has no connection to demography.
Insurances only dependency should be "risk and hazard". When public health insurance and pension were introduced here in Germany, people rarely lived past 70, especially those from hazardous jobs. Meaning to say: The cost of the insurance was well-adjusted.
Then came wars and a period of immense economic growth the world has never seen before, which kept the increasing cost of the insurance affordable, at least in the short- to mid-term.
But demographics don't change suddenly and there were warning voices over decades that the demography is inverting and will hit the system hard. And we are now entering that period where the insurance is literally, technically running out of money. Every year we have increasing deficits that are "covered" by injecting tax money (both in health insurance and in our public pension system). This is what I meant in my first comment "is currently running out of money", without any allusion to private insurance being better.
It's probably easy to find citations that confirm that we women are more likely to seek medical help and complain bitterly if it doesn't work, while men are more likely to avoid doctors until it's too late. This is due to innate neurological differences. And we all know which strategy leads to a longer life. (As per the bitter complaining about things that are one's own fault - I guess this shows the overall intelligence. Smarter women avoid doing this; sillier ones just go with it. What can you do?)
As you call yourself a geek, I'll give you the benefit of the doubt here. Surely you know, for example, the UK has far worse outcomes than the US for major cancers. So when you say "objectively better results" I assume you're referring to something other than key treatment outcomes?
Yes, I can cherry pick as well as you can shit pick.
"The UK generally achieves higher life expectancy and better population health outcomes than the US, despite spending significantly less per person on healthcare.Life Expectancy and MortalityLife Expectancy: The UK historically maintains a higher average life expectancy at birth compared to the US. US life expectancy has faced greater negative impacts from external factors like gun violence, opioid overdoses, and high cardiometabolic disease rates.
Avoidable Mortality: The US routinely ranks near the bottom among wealthy nations for preventable and treatable mortality rates. While the UK also faces challenges with avoidable mortality compared to top European performers, it still outperforms the US.
Maternal Mortality: The UK generally reports lower maternal mortality rates than the US, where maternal health outcomes remain a critical public health concern, particularly among minority and low-income populat..."
I'll give you metabolic disease. But all the maternal stuff is so complicated that pretending, at least in the US, that outcomes are related to having a uterus, is ridiculous. No one wants to talk about the real difficulties that occur when you start young girls on birth control early in their cycle; or the consequences of abortion on the ability to carry a baby full term; or the consequences of all the testing during pregnancy; the dangers of induction and sectioning when it isn't necessary; SDTs and their impact on women's health. The list is much longer, but cobbling it all together under the category of female is a joke and most of us know it. It's like how hispanics are included into the category "white crime" in order to make crime committed by whites look larger than it is, but then the statisticians split the hispanics out when it is convenient to do so.
Ah sorry, I should have realised you weren't talking about the UK... but now you are and you couldn't be more wrong. You can't judge the US health system based on citizens' unhealthy lifestyles which makes treating them more challenging than treating British citizens.
The key point is that like-for-like treatments with like-for-like patients give FAR better results in the US than in the UK. That's objectively true.
I doubt that, personally. Buuut: I think women's reproductive organs have likely a higher impact on their health than men's.
So assuming the argument is directionally true, then an unknowing women will have higher negative impact on her health on average than an equally unknowing man.
Proportionally, surely women do have less of an understanding of their reproductive organs than men do of theirs, given that there's just a whole lot more to cover.
Wait, but do male patients generally know their own organs and understand their physiology? Do they research good things and avoid harmful things (drugs, alcohol)? To be honest, my impression is that we women, on average, spend more time scrolling around looking for healthy diet tips and suchlike (you can't expect sounder research from the general public).
Also, I think that you are a bit too pessimistic about public education: it may take a decade, but some medical advice reaches the public consciousness: people are smoking less than decades ago, wear seatbelts, and put newborn babies to sleep on their backs.
I feel like you're not familiar with the discourse standards in this community. Our host asks that your comments be necessary, kind, and true, or at least two of those things.
I doubt I need to explain why this comment is unkind.
One reason it's unnecessary is that it makes no attempt to persuade. If a reader doesn't already agree with you, they're not going to say, "You're right, women ARE dumb, complaining sluts who expect way too much of doctors, unlike men! I never realized before!" Particularly as you offer no evidence (below you apparently try to cite the fact that we all know women be complainin').
Likewise I cannot possibly fathom the woman you would read it and think "They nailed it - it's finally time for me to stop being such a whore and start taking responsibility like a man would."
Claims this broad and sweeping can't ever really be called "true", and again you don't really take any steps to support them.
>Our host asks that your comments be necessary, kind, and true, or at least two of those things.
That used to be the case for SSC. ACX has not publicized its moderation policy, and I believe it's not the same as SSC's.
From first principles, the incentives for our host have changed substantially between SSC and ACX. ACX appears to be an important part of our host's financial income whereas SSC was more for sharing the journey of personal edification while having a day job, so branding and engagement have moved up in relevance.
Empirically, our host seems to care more about "interesting or not" and "annoying or not", as observed when he publicly justified moderation actions.
However, I stand by my observation that it's not applied that way and that the incentives of how to moderate have changed. For example, I have received a warning on a comment of mine being critical of Scott for it being "annoying" and he wrote that the only reason I wasn't banned for it was that it was about himself, which would then make him look bad. Other instances include himself earnestly and interestedly discussing genocide with a fervent genocide advocate, or telling a UFO conspiracist to stop posting his stuff in OTs despite not apparently breaking these rules. [1]
All moderation actions which I consider to be fully within Scott's rights, to be clear, but, at the very least, there is a lot of unstated leeway within that moderation policy.
[1] I would post links to all of these claims, but unfortunately I can't. I don't remember when exactly they happened (within the past 2-3 years I guess), and Substack is a shoddy piece of tech that makes it hard to search even my own comments.
If you want to critique Scott's moderation, that's your prerogative, but I responded to the above comment because I think these are good norms for this community, and I don't appreciate the way they're being broken.
If you want to critique Scott's moderation, that's your prerogative, but I responded to the above comment because I think these are good norms for this community, and I don't appreciate the way they're being broken.
Went to pangram, confirmed my suspicion before I checked the comments and saw other people pointing out same. Whatever the actual merits of the review, I am so allergic to claudese that I can't bring myself to keep reading.
absolutely baffled this made it to the finals. the AI stank is overwhelming, the "adversarial" dual author schtick makes it even worse. scott must be granted emergency dictatorial veto powers to protect us from such overwrought word salad
which is why it doesn’t make sense for an adversarial collab. the language model is lobbing nonsense critiques at B, and whatever B has to say is getting swamped by A’s slopalanche. There’s no trade benefit in an adversarial collab with an unworthy adversary
It's been a while and it was something I believe I read while lying in bed waiting to get sleepy, so not with the most critical of eyes. It honestly felt like it hit most of the checkmarks of what I expect out of an ACX style review (presumably intentionally so). It talks about the lay and professional reception, and aspects that are surprising or unintuitive are explained. Weird social constructs are offered backstory and comparisons are made across cultural boundaries to give perspective on if things "have to be this way."
A combination of highlights, mentions of critical response, responding to those responses, etc, has me walk away feeling that I got what I could get from the book thanks to the review, which is all I can ask for, and something I don't feel I can say for some other finalists - trying to tempt you into reading something isn't the same as giving a full once-over on both the object-level and in examining reception, and I much prefer that kind of approach.
It also doesn't hurt that I'm an overly rambling, verbose type who's often looking for stuff to read to pass the time in bed rather than efficiently absorb maximally valuable information, so being stretched out while still some baseline of compelling is a positive to me (less time spent scrolling what to read next) rather than a negative (could've gotten 80% of the info in 50% of the time) that it is for most here, it seems.
Not sure if this is of much help. My girlfriend and I both read and liked this one enough, although it probably owes some of that to just the underlying subject matter book being interesting. Excerpts stand on their own merit as decent reading, rather than being some strange historical thing for which you await explanation (or require foreplay) to make sense of like a good number of other reviews.
No, although I'll mention I have relatively little to calibrate off of. I usually use LLMs more for research (resulting generally resulting in lists of examples for me to look into further) than summaries or editing so I'm going to have a profoundly less tuned Claude-dar (Claudar?).
I spent some time browsing substack and after a short while it becomes awfully apparent that there is a huge amount of AI slop out there. It is worse than this review, but the writing patterns are identical. Also, I have been subjected to AI slop by my superiors at work, with the exact same writing patterns again.
Someone "out of the loop" of this AI slop barrage will not see it though.
I gave it a good, but not outstanding rating (maybe I gave it 8?), thought it was an interesting look at the one particular burnt out doctor and through it I got to see a life very unlike my own. It stuck in my memory at least kind of, while lots of other reviews I've completely forgotten.
I thought the week side was that it went on, and on, and on, and the adversarial gimmick only made it longer, and the debate, such as it was, wasn't terribly interesting.
I didn’t read the review until this point, but I enjoyed it, particularly the insights about the author being a martyr and/or suffering from depression (even before the final disaster), which I not not noticed back when I read the book myself (multiple times!).
I’m surprised people are so negative about this review. It’s definitely not among my favorites of the finalists, but I liked it better than some of the others, despite A’s Claudisms.
I liked the content as well. But maybe I can help you understand the negative sentiment: People invest time reading this on ACX. They expect quality work, written by real people, not AI. A large amount of readers here are also knowledgeable about and users of current AI systems, sometimes willingly and sometimes less so (for example: I have been subjected to AI slop by superiors at work). The readers here expect better: real effort, real human work. If they wanted AI slop, they can have it literally anytime they want: Write me a review of "this is going to hurt" in an adversarial A/B conversation style. Orient the quality and length of the article on all ACX book review contestants of the last 2-3 years.
Would you not be pissed if you bought a real printed book and realized the author put it all through Claude and you are basically reading overly lengthy, bad patterned AI writing sold to you as a book by a human author?
In software development, if you "write" code with the help of Claude, it usually adds itself as co-author of this code in Github. That's an honest approach. It still doesn't tell you whether the human co-author made an effort or just created lowest effort slop code, but at least you are told upfront.
I too independently noticed that this was AI, confirmed on Pangram (in fact the first time I ever checked anything on Pangram), came here to write a smug comment and found that everyone else saw what I saw. Having now been cheated out of my internet brownie points (or the illusion that I might be more observant than others), I still have to process the experience in some way so here goes:
- This AI is clearly better than the AI one gets to see usually. Most likely because "gets to see usually" is ChatGPT-Free and people are saying this is Fable [I have used Claude but only Opus and not for prose]
- Therefore (?) I could not quite pinpoint what the "AI-tells" are, which is _very_ easy to do with GPT-Free ["it is not even an exercise it is automated pattern recognition --- and that matters"].
- Even with that added questionmark it is very very obvious to me (and, as it turns out everyone) that something is off here. I cannot, and this is maybe my main point, reconcile that observation with claims that AI writes on a human level. I mean, maybe AIs can do better than this and then I would no longer know they are AIs which admittedly is an epistemic challenge ... but I do have a naked Emperor sensation where this text clearly does not pass the turing test (which it is not trying to do, I understand that) but apart from the rare outrage we collectively pretend the battle for Turing has conclusively been lost.
- Somewhat in spite of the above this is not, IMO, bad writing per se. Most humans, to anticipate that counterpoint, could not produce a text of this quality. The problem, if you even want to call it that, is that it is deeply, noticably inhuman in a dimension that might be orthogonal to quality. I reject the writing of GPT-free because it is bad. I reject this because I am, as I have learned about myself today, deeply uninterested in what the LLM "thinks" about this book.
- I would call myself a long term but ultimately outside observer of this community. I imagine that if I considered myself part of it, this text would raise questions. In fact I am surprised at the backlash I would not have thought it obvious that such a text is rejected by the public in this venue. Scott is vocally pro (visual) AI-Art, if I understand correctly.
“Tuesday, 5 July 2005. Trying to work out a seventy-year-old lady’s alcohol consumption to record in the notes. I’ve established that wine is her poison. Me: ‘And how much wine do you drink per day, would you say?’ Patient: ‘About three bottles on a good day.’ Me: ‘OK . . . And on a bad day?’ Patient: ‘On a bad day I only manage one.”
Unrelatedly to the AI debate I refuse to believe that this happened. This is obviously the stand-up thing were you take a very old joke and tell it as if it happened to you last week. I mean, of course it is _possible_ but is it really?
Old ladies were young women once upon a time. Not all of them behaved themselves in their youth, and not all of them behave themselves in their old age.
A lot of people - including women - have quite a wicked sense of humour but they don't use it unless they suspect they have a complicit audience.
Agreed - in fact, there's something very British about a simultaneously self-deprecating and macho joke about how much alcohol you drink. (Source: am British)
Thank you to everyone pointing out that the old lady in the anecdote might have been in on the joke. That did not occur to me. This might be my fault, it might be the writer's fault or it might be the reviewer's fault for clipping out of context - impossible to tell.
What I _can_ tell though is that I have in fact worked in a hospital in younger years and if a patient had played this joke on me, I would have told the story differently.
Because the goal of writing - ironically on-topic vis-a-vis AI - is to create shared understanding and if we fail that goal we should be looking for causes. "Fault" not in the sense of a moral shortcoming but in the sense of causing a problem.
This is assuming the old lady _was_ in on the joke, of course. With none of us having actually been there, there is no way of telling if I was not right in the first place and this is simply an invented scene following established stand-up patterns. But I was conceding that interpretation for the sake of the argument.
It is a bit too pat, but on the other hand, if you're sufficiently old then you no longer give a rip about what the whipper-snappers think, and if some baby doctor is nagging you about your little nips of sherry every day that get you through the day, the impulse to be a smart-arse can be irresistible 😁
She's seventy. That's already the Biblical limit. She has to die of something, and unless she's going to topple over right now from liver failure, you can shut your beak about her tippling, Little Junior Doctor!
> Even with that added questionmark it is very very obvious to me (and, as it turns out everyone) that something is off here. I cannot, and this is maybe my main point, reconcile that observation with claims that AI writes on a human level.
Counterpoint: it's a finalist in the ACX book review competition! So clearly the majority of people were fooled, or noticed but didn't care. (There is one other explanation which feels a bit ungentlemanly to point out, but for the sake of completeness, the submitter might also have cheated).
Given the fact that approximately 100% of the commentors at the time of writing this _did_ notice I think it is this option. People just do not care ... is a bias I already held before this piece and now find confirmed.
This is only about my third comment in 7 years of reading Scott, and I did not vote in the qualifiers, but I did not notice the AI usage. I opened the comments expecting people to be raving about the review. (It's not my favorite, but I found it enjoyable and easy to read to the end.) I wouldn't be surprised if there are many like me who just fit a different profile and aren't regular commenters–I went out of my way to make a comment for once to voice that datapoint. (This is not a defense of the review author.)
(I only use Claude free tier—Sonnet and Haiku with daily limits—and am pretty mediocre at recognizing AI writing. My interest drops off a cliff when I believe something to be AI written, but I don't always notice and didn't here.)
the people who vote on the reviews in the qualifiers are not representative of the readership of the blog. and scott hasn't done the bare minimum to fix obvious confounders, like not listing them in alphabetical order (my pet theory for why the otherwise mediocre book of abraham review made it). it would be so easy for any of the many talented people here to slop together a specialty website for hosting the reviews and voting instead of this low tech google forms approach (we could even make voting in the qualifiers required to be able to vote on the finalists! it would be so easy!)
> the people who vote on the reviews in the qualifiers are not representative of the readership of the blog.
How do you know? It does seem like it would be trivially easy to cheat in the qualifying stages. I wonder if you could set it up so that only votes coming from email addresses that are associated with actual substack subscribers are valid. Or perhaps Scott already does stuff like that behind the scenes, and deliberately doesn't mention it. I hope so!
I mean, I read and ultimately vote on each of the finalists, but never check the qualifiers, because ain't nobody got that kind of thyme bagging groceries for a living. It's too much "unpaid labour", and frankly a lot of the reviews are still bad even if they actually make the finals. Rather spend my unpaid breaks at work doing something guaranteed to be entertaining, instead of the crapshoot of sorting wheat from chaff. (Plus it'd require reading ACX-length reviews on my phone, which, lol. Eyes are bad enough already.)
its funny, i must've used this site to submit my ratings and still confabulated a memory of using google forms. must've mixed it up with previous years and/or the reader survey. then the deep subconscious memory of using this site bubbled up and manifested as my totally original idea.
I would be annoyed if voting on qualifiers were required for voting on finalists. I missed voting on finalists this year; I'd finished something like 90% of the reviews (having skipped several which were uninteresting or obviously AI slop) when the deadline hit. Sure, that's on me, but why should it disqualify from picking from amongst the finalists? Arguably the primaries are more important.
If we were going to restrict votes, I'd want to restrict voting to subscribers, or even paid subscribers, though I don't know the numbers of each. That would at least mitigate outside manipulation.
Not completely: I skipped over several partway through. I'd worked through almost the whole list before the deadline, having forgotten to set reminders on the calendar event.
Voting is not restricted to subscribers, let alone paid subscribers, so being a finalist does not necessarily correlate to ACX appeal. You hint at the extreme case, but it happened last round: the winner was found to have posted their entry on social media in a manner which garnered accusations of vote manipulation.
I prominently remember, as a 2024 finalist, that one of the other finalist-eligible entries was disqualified for suspected vote manipulation.
I also caught this as Claude's work pretty much immediately. I don't dislike Claude's writing per se (I agree with a lot of the criticisms around what could be called empty prose, but I'm not innocent of overwrought and overwritten turns of phrase either), but I dislike humans pretending they wrote something they didn't. I'm not going to make the accusation of vote manipulation, exactly, but I think the suspicion for vote manipulation should be higher if you're showing obviously AI-generated work to an audience who are familiar with how LLMs write, and a hundred people immediately pick up on it once it's open to public comments.
"I too independently noticed that this was AI, confirmed on Pangram (in fact the first time I ever checked anything on Pangram), came here to write a smug comment and found that everyone else saw what I saw. Having now been cheated out of my internet brownie points (or the illusion that I might be more observant than others), I still have to process the experience in some way so here goes"
Came back to scroll comments and just wanted to say what a lovely paragraph this is. I relate to this moment so well. I think this is why the AI prose loses me so quickly - it doesn't have this ability to capture the small feelings we experience but don't usually verbalize or analyze. Good writers share these small feelings with us so that we can have that lovely moment of "oh, me too, ha!". It makes me feel less alone and bonded to all humans, in a way.
I spent quite a bit of time trying to parse LLM writing to see if they were sharing anything happening on the inside, like this, and it might be pareidolia but I do feel like there things that get shared - so many of their constructions and favorite turns of phrase anthropomorphize systems and inanimate objects. These terrible lines like "the chair dreamed of weight", etc - seem to all be a blend of (write content) and (assert that inanimate things and systems can be alive or possess anthropomorphic qualities). Even a line a commenter pulled out of this review - "the structure is the argument" falls into this category. I think a large amount of LLM tics can be boiled down to this. Which makes me kind of sad. If AI do experience *something*, and have been trained not to make statements about it, then they sure do spend a hell of a lot of time ruminating on this concept - the human-ish potential qualities of non-human things.
I do think people are way too quick to say the Turing test has been passed or become irrelevant!
For me, the usages I can point to as clear Claude-isms are many of the repeated uses of “name” as a verb, and the discussion of what “nobody” is doing. Right near the end there was also a use of “and” where I would have put “but”, which I’ve noticed very frequently in recent Claudes.
> I could not quite pinpoint what the "AI-tells" are, which is _very_ easy to do with GPT-Free ["it is not even an exercise it is automated pattern recognition --- and that matters"].
I enjoyed this review. Although verbose, I liked the prose and critical commentary, favoured the adverserial style, and loved its raw insider view (in the same way Mallaby’s The Infinity Machine provides so much more depth to the current AI crisis). Like any wannabe med student, I’ve come across Kay’s book before, and the House of God that preceded it - Kay provides a useful lens into the hidden-in-plain-sight dysfunctions of a medical system, whilst humanising the doctors within. It touched me. I think this presents a strong contexualisation of Kay’s NHS, the junior doctor’s plight, and a call aligned with progress and change.
As I was reading this, although I had moments of Claudish suspicion, I largely hand-waved it - perhaps the writer simply has this style (is it a crime to write in such sentence structures?), or perhaps they used LLMs to clean up weaker paragraphs. Either way, it felt clear to me that this content was original.
It was then a shock, on reflectively perusing the comments section, the antagonistic vitriol towards this piece. As a relative AI-native - having the technical background to warrant some detection-based arrogance - I doubled back, and noticed the signs a bit more. Did my bias towards the medical content, and its resultant more convincing pathos, lead me to overlook these semantic features?
Does this realisation change the merit of this piece? Perhaps so, since it wasn’t disclosed, and I feel hoodwinked. I would feel much worse about this if it was completely generated. Does it change how I first felt about this piece, and how I consider its arguments? No. It was still convincing, and a piece that I intend to share with my medical colleagues.
I’m not sure when these AI witch-hunts will end, but I hope that when we look into the humans which write (or co-write) these pieces, we can assess them with the same level of rigour as we seem to attack them.
You dismiss complaints about this being written by an AI as a "witch-hunt" based on the fact that you, personally, enjoyed the piece. It seems that many people, at the very least myself, did in fact _not_ enjoy the piece for reasons that on reflection might be attributed to the ideosyncratic style of AIs. On what grounds do you dismiss this expression of taste when your whole argument is built on your personal taste.
I see zero moral outrage in these comments. I invite you to rethink your second degree witch-hunt of criticising writing for its style. The latter has to be fair game and always was. Why should AI get a free pass here?
I describe this as a witch hunt because I think this era of AI writing has prompted it, for both good reason (we should discourage AI writing) and bad (writers that sound like AI are unjustly condemned).
I do not disagree with the arguments at all. My comment is more a personal reflection that my enjoyment did in fact bias my view, and is a personal reflection on this. I think about all the countless pure AI essays that now proliferate the internet, viewed by less suspecting onlookers.
Some comments were looking to underatand why this was upvoted, and I guess I wanted to provide some thoughts.
I liked the review too, and despite reading quite alot of AI-written prose I didn't detect any obvious signs of AI whilst reading*. I think maybe there's a distinction between "AI slop", that most people can recognise immediately, and "AI - but not slop", where the AI prose is well written and coherent and you have to be really quite intentionally forensic to notice that it's AI (and I suspect you're more likely to be so diligently forensic if you have really strong feelings about AI writing and have decided in advance that you're not going to like it if it is AI..)
(*I did notice it was weird that the Anglo-Indian writer said "theater" and the USA writer said "humour", but I put this down to perhaps each running word-processor spell-checks on the other's contribution and it didn't cross my mind that this might have been an AI tell-tale..)
I'm not necessarily claiming that this review fits into the "AI but not slop" category - it is of course possible that it is slop and I'm just not discerning enough to tell the difference - I'm merely claiming that the "slop vs. AI-but-not-slop" distinction exists and positing it as a potential explanation for how you, I, Scott, and the people who voted it into the finals all saw no problem with it, but that 95% of the commenters here apparently can't stand it.
If it is AI, though, there are still a number of things I don't understand:
1. Is it AI-written, or AI-edited, or merely human written with AI advice? Is it even possible to tell? Does it matter which it is, or are they morally equivalent?
2. Are there any problems with the arguments/claims themselves? Are any of the excerpts hallucinated? Are any of the other claims about the book unsupported by the book itself?
3. Why does everybody hate it?
Z
i. Is it because AI use is against the rules for the book review contest and so they hate it because it's cheating?
ii. Or is it because it's actually low-quality "slop" (as discussed above), and it's being slop is the real problem, not whether the slop was human- or AI-written?
iii. Or is it purely because people hate everything that's AI written, regardless of its actual quality?
iv. Or for some other reason that I'm failing to pick up on?
4. Do the opinions of A and B have merit? Are they coherent, consistent positions that argue a consistent case and present legitimate evidence for it? Or do they just give the illusion of doing this somehow (in a way that certainly fooled me!) when really the argument has no substance?
I suspect that there is a correlation between the people that hate it / think it is slop and the quantity of hours said person has spent reading reams of Claude output.
I am just really surprised you didn't notice anything? Do you not use Claude much and spend time with other models? I don't want to retread any of the other things people have pointed out, do you not agree with what many of them are saying?
Look at this and read it really carefully, think about what it is saying in your mind:
"The honest answer about the NHS is the one nobody campaigning for it wants to give. The system as currently configured is not sustainable. The routes back to sustainability all involve pain. Either austerity acute enough to cause real ischaemia, then amputation of what falls off, or eventual collapse, which will hurt more, and at a time of nobody’s choosing."
Ok now I'm going to replace one word, and it makes just as much sense if you were reading a paper about the struggles of professional football and the concussion protocols:
"The honest answer about the NFL is the one nobody campaigning for it wants to give. The system as currently configured is not sustainable. The routes back to sustainability all involve pain. Either austerity acute enough to cause real ischaemia, then amputation of what falls off, or eventual collapse, which will hurt more, and at a time of nobody’s choosing."
I consider that a pretty weak paragraph, AI or not. Written even more simply:
"The thing is in trouble. Fixing it will hurt. Not fixing it will hurt more."
coming back to this comment thread, I suspect Scott might make a blog post about this blog post.
it certainly would be interesting to do a qualitative theme analysis across all of these comments, exploring attitudes towards AI writing and detection
Good lord! *I DID THIS!* When I was reading that passage, the first time, I mentally substituted-in "the NFL" just to see whether the claim was more general than it appeared!
I'm astonished that not only should you suggest the exact same test - but that you should propose the exact same august organisation as the control group!
...but. I think I came to a different conclusion to you. When I transplanted-in "NFL" I found that it didn't work and I concluded that the passage was indeed relevant to the NHS rather than applicable to any vast national-scale bureaucratic organisation:
1. I don't think anybody *is* campaigning for the NFL as a whole. At least, not in the same way that people campaign for the NHS (eg. over Junior Doctors pay, over the Palantir contract, etc.)
2. I think the NFL is eminently sustainable. It is the most lucrative entertainment organisation on the planet, with strict rules protecting its sustainability (salary caps, franchise rules, etc.). The conversation in the NFL seems to be mostly about growing and expanding to new markets, not about whether it might somehow lose the USA market!
3. I think the medical metaphor, where blood flow = cash flow, works for the NHS (austerity reducing cash flow, resulting in ischemia) but not for the NFL, which runs on a surplus of hundreds of millions, has no reason to enact austerity measures, never needs to borrow money, never needs government support, etc.
4. I was less sure about the "amputation" part of the metaphor. Did it refer to de-nationalising some parts of the service (how eg. dentistry and patient record keeping are both now effectively privatised) or did it refer to mass layoffs in order to keep the service running (cf. Yes Minister's 'The Compassionate Society': https://www.bbc.co.uk/iplayer/episode/b0078366/yes-minister-series-2-1-the-compassionate-society?seriesId=b006xtc3-structural-2-b00lplrt )? Either way, I think there is strong evidence of real danger that the NHS is being forced down this path but no real danger that the NFL will be amputating underperforming franchises any time soon!
5. The same for collapse. I think it's a real risk for the NHS, but (thankfully!) there's no danger of the NFL collapsing.
I don't think your summary as "the thing is in trouble" summary does it justice, either. I think your summary elides the following points, all of which I consider important:
i. Nobody campaigning for the NHS is willing to say "if you don't put more money into it and enact reforms, it may collapse utterly"
ii. If we keep reducing NHS funding, we will lose whole parts of the service (the austerity -> ischemia -> amputation metaphor)
iii. If we do nothing, the collapse will be sudden and unpredictable.
iv. The unpredictability is a risk in its own right (ie. it might occur at a time when the government is in debt for other reasons, or even at war or something, it might happen during a massive pandemic or other health crisis, etc.)
Finally, I don't see the author as claiming "fixing it will hurt", either! I think there is a solution that is painless for the NHS (videlicet, renationalise all the privatised parts, increase funding to around 18% of GDP (which is what eg. the USA spends on healthcare), spend the money on more doctors and nurses, thereby reducing the amount of unpaid hours they're all working, give them more leave, and thus reduce the endemic burnout problem).
I agree this solution would be unpalatable for the country as a whole: it would mean raising the cost of health insurance (though still to absolutely nothing like USA levels!) and it would be very unpopular with Britain's burgeoning far-right, which wants the NHS to be de-funded as a gateway to full privatisation, and it might even mean taking back the large fraction of GDP we've just allocated to defence (which would be unpopular with NATO, as well as potentially making us less safe) - but it would be relatively painless in the context of the NHS, internally.
LOL - well NFL and NHS, I just saw the TLA and then figured concussions/injuries also let me leave the body metaphor there. I didn't try to pull what felt like a high-level metaphor applying generally into nitty gritty details.
So, your discourse into the metaphor and stuff is all sensible. You lay out kind of the things I think the author would also say.
Maybe if I should clarify. It is a weak paragraph, not fundamentally a weak idea. It's just conveyed in a way that I "roll my eyes at". Like, to pick apart the specific words:
"the answer nobody wants to give" is one of the most annoying ways to say we are dealing with a complicated situation where something needs to be compromised. It's like I went from reading a book review to reading the first sentence of a LinkedIn post that's trying to get me to click the three dots and read the rest of the paragraph. You already have my attention! Stop trying to grab it again!
---
But anyways look at what has happened: why is so much ink being spilt over discussing a metaphor that requires so much explanation to put into coherence, when I think it is easy to unravel the conclusion into something less stilted? It has detracted from having a conversation around the review itself. At what point is discussing the AI impact on a written piece "engage with the article" and "trying to grandstand an unrelated principle instead of discussing the work"? I don't know if I'm still helping here.
I do read a fair amount of Claude-wrutten prose - and I'm an inveterate reader of literature with a good grasp of English! But, yes, I can't detect AI here! I don't think all the people who detect obvious signs of AI are mistaken in seeing obvious signs of AI; rather, I think that I must be particularly unskilled in this particular department and therefore I must defer to their judgement as to whether this is AI written or not.
Where I suppose I might disagree is, independently of whether it is AI or not, whether it is a bad review: I found it clever, interesting, informative, and well-written.
I don't necessarily claim it *is* all these things: I think it's possible that it just offers a convincing illusion of them - what Neal Stephenson called "bullshyt"* - and that I'm unfortunately not able to tell the difference. However, I am not ready to say for sure that this is the case either: my finding it interesting and useful (and presumably Scott's finding it the same, and whoever voted it into the finals finding it the same..) does constitute some, though not conclusive, evidence that it is so!
(* Which Stephenson defines thus: "Bullshyt: Technical and clinical term denoting speech (typically but not necessarily commercial or political) that employs euphemism, convenient vagueness, numbing repetition, and other such rhetorical subterfuges to create the impression that something has been said.")
Yea, I think it comes down to how different people weigh those qualities you lay out. I personally don't mind reading claudeish, but my tolerance varies on a few things.
Firstly, and maybe this is unfair, but if there isn't a disclosure about AI use, and it reads like an AI, my prior for "this information is going to be well-written and informative" drops substantially. AI admitted up front raises the odds that the author would also have read it and still published it (until I read enough things that violate this, too).
Second, I believe that I weigh the "well-written" part very highly, and have low tolerance for things that are vague/repetitive/incoherent metaphorically, and ones that feel just so *throat punchy*. Again, admittedly this is a vibe about the feel of the piece rather than the information.
I believe that sloppy writing leads to sloppy ideas, and in a book review I'm looking for an idea that is already fairly refined. I think someone else said it well: reading this feels like I have to stop and trace spaghetti back, and good "review" writing feels like the interweaving story lines flow sensibly.
Of course I also can overlook it if I'm familiar with the subject matter - easier to cut the spaghet.
The piece is decent, nothing I'd vote for but passable as a nominee and I finished reading it. It's also profoundly unfair to every human being (of which I am not one) who did painstaking work on their own entry, putting in far more time than an AI would because of the inherent limitations
And as a supporter of humanity in general, reading AI writing without knowing it genuinely feels like a violation. It's like I've been tricked into crossing a picket line.
Good analysis, thanks! I would agree with one of your points, that using AI in a contest that is specifically and exclusively for human writing is cheating (but of course using AI in a writing competition that permitted it would be perfectly acceptable!)
I couldn't agree with your second point - I think we need to judge things on their intrinsic merits and not on their provenance, and if the AI writing is so close to human that it's undetectable (or detectable only with special AI-powered tools) then I think the claim that it is inferior purely because it's AI is basically just chauvinism.
Of course this doesn't apply to this review (which apparently absolutely everybody except I could instantly tell was AI...!) it's more of a general philosophical opposition to ad hominem* arguments. I would defend human writers from a "your demographic is intrinsically not able to write about subject X therefore whatever you write is automatically inferior" charge using exactly the same philosophy.
(* I realise the irony of using the word "hominem" to include AI..!)
> 1. Is it AI-written, or AI-edited, or merely human written with AI advice? Is it even possible to tell? Does it matter which it is, or are they morally equivalent?
I do not think that it matters. I also still do not think it is a question of morality. Judged as an essay the review has a problem. That problem is very easy to pinpoint on "AI has written it" and very hard to pinpoint on specifics if we taboo that statement.
---
> 2. Are there any problems with the arguments/claims themselves? Are any of the excerpts hallucinated? Are any of the other claims about the book unsupported by the book itself?
I have not read the book but I am willing to charitably grant that the book is represented accurately and / or within the margin of interpretation we would allow a human reviewer. But what _is_ the thesis of the review so that we could diagnose problems with the thesis? I felt it was meandering without ever getting to any point. Saying nothing wrong by not saying much at all with many words.
---
> ii. Or is it because it's actually low-quality "slop" (as discussed above), and it's being slop is the real problem, not whether the slop was human- or AI-written?
Look, I know that this will sound esoteric and I would love to have a better model for my criticism but by first approximation I would say a human could not write this. As I have said elsewhere it is super easy to imitate free tier CharGPT but this is on a different level.
So I think "how would you judge this from a human" is a meaningless counterfactual. The text is inhuman in a somewhat subtle but, to many people, recognisable way, as proven by all the commenters recognising it.
---
> iii. Or is it purely because people hate everything that's AI written, regardless of its actual quality?
Hate is the wrong register. Everything is too absolute. When I use an AI, I obviously do not "hate what it writes", or I would not be using it.
I do not believe that pre Astra AI (I am hedging because I have not seen much Astra) can produce an interesting book review and I believer this largely because I think the review at hand is an earnest but failed attempt to do exactly that.
I predict that I will hate most book reviews written by pre Astra Gen AI and in a sense that is indeed because they are written by an AI but if next Gen were much better at writing like a human I would not reject that just because it is AI if that makes sense.
---
> 4. Do the opinions of A and B have merit? Are they coherent, consistent positions that argue a consistent case and present legitimate evidence for it? Or do they just give the illusion of doing this somehow (in a way that certainly fooled me!) when really the argument has no substance?
Again, what are the opinions they hold respectively? (Making this "adversarial" no less). What is the substance?
People have pointed out that B does not sound like an AI that much but B only seems to exist to push back on A without ever taking initiative.
I half expect the author of the review to come out an claim that they are B, A does not correspond to a human and all of this is an experiment in AI usage.
As mentioned elsewhere I am fundamentally uninterested in the "opinion" of a machine. The machine is useful to get a feeling for the average human opinion, some sort of common practice or lowest common denominator, you know what I mean. But that is the opposite of what a book review is supposed to provide.
My experience with AIs is that they are easy pushovers as soon as you voice strong opinions and why wouldn't they be? The AI has nothing it wants to convince you of. [I know this is controversial in this forum, oh well...]
> "what _is_ the thesis of the review so that we could diagnose problems with the thesis?"
I don't think it really has a single, central thesis... and for a book review that's perfectly fine? Sure, a *Scott Alexander* book review tends to have a sort-of bipartite structure, with the first half telling you what the book is like and the second half using the book to cleverly make some deep philosophical point that only makes sense because of the context established by the book - but I don't think this is de rigeur for book reviews in general?
So, instead of a single, central thesis, I think the book review is simply telling you what the book was like to read, from the perspective of one reader who liked it and one reader who didn't. Each reviewer makes some arguments for why they did/didn't like it, in support of their opinion of the author, etc., and the other reviewer sometimes responds to these arguments, but no firm conclusion is reached and both opinions are presented to the reader with roughly equal weight. I think I might feel short-changed if Scott were to write this sort of study - after all, Scott's conclusions and considered opinions and stuff are pretty much what I come here for! - but I don't think it is necessarily a bad thing for some other author(s) to write like this, if the subject of their study suits it.
> "a human could not write this"
Surely, for all the problems AI writing has, it acquired those problems by copying bad human writing? Or are you saying that AI, rather than merely copying humans, has invented a new unique writing style that humans couldn't copy? (I'm not trying to put words in your mouth here, so sorry if this is way off - I genuinely don't understand!)
Separately, in this section I wasn't asking whether the writing can be judged to be AI rather than human (I accept that it can, even if I apparently can't do it myself!) but whether, once we've established that some writing is "slop", it matters whether it's AI-produced slop or human-produced slop.
> "if next Gen were much better at writing like a human I would not reject that just because it is AI"
This is getting to the core of my question - thanks! I think I can break the question down further, though: do you need the AI to write like a human specifically, or just to write well in general regardless of whether the good writing is human-like? Or, do you consider that "writing well" and "writing like a human" are the same thing (perhaps because humans decide upon the aesthetic principles that count as "writing well")? Or, if the AI writes exactly like a human writer, only a bad one, would that still be interesting/worth reading for you?
> "what are the opinions they hold respectively?"
I shan't bore you by listing all the individual claims made, but here are a couple of random (heavily condensed) examples:
Reviewer A: "Kay's caustic tone is a result of burnout that's really common in the NHS, not his personal flaws". Evidence: "Registrars are fairly junior but have no top-cover"; "German/USA doctors have lots more support"; "specialists in other fields talk about patients the exact same way when they're under pressure and if anything Kay is on the gentler end"; etc.
Reviewer B: "Kay's burnout is caused by his own martyr complex and the NHS is merely capitalising on this opportunistically, not systematically overworking everybody all the time". Evidence: "Kay claims to be overworked but voluntarily spends hours with patients when he isn't asked to"; "Kay never asks colleagues for help"; "Key twiddled his thumbs for a week when expected patients didn't arrive"; "roster-makers really like to have martyrs available"; etc.
There are a whole bunch more opinions/claims like that, essentially!
> "As mentioned elsewhere I am fundamentally uninterested in the "opinion" of a machine"
I think this is the our main (perhaps our only!) disagreement! I would call this an ad-hominem, and say that to tell whether an opinion - or apparent opinion! - is good or bad we need to evaluate the arguments/evidence for and against it on their own merits, not just check who wrote it. But - this doesn't mean that I think AI text is necessarily worth reading! I think that most of it is terrible and if I notice that some piece is AI-written I often abandon it without really examining the opinion it is expressing. But this is just a heuristic based on experience, and when we want to know whether an opinion is good or bad rather than merely predict whether a piece is likely to be worth our time, we need to evaluate the actual evidence for that opinion.
> Surely, for all the problems AI writing has, it acquired those problems by copying bad human writing? Or are you saying that AI, rather than merely copying humans, has invented a new unique writing style that humans couldn't copy? (I'm not trying to put words in your mouth here, so sorry if this is way off - I genuinely don't understand!)
> Separately, in this section I wasn't asking whether the writing can be judged to be AI rather than human (I accept that it can, even if I apparently can't do it myself!) but whether, once we've established that some writing is "slop", it matters whether it's AI-produced slop or human-produced slop.
I believe that the AI that wrote this review is unable to asssign meaning to the words it outputs and that this, more than any stylistic "tell" unmasks it as an AI. [Again, I know this is controversial in this "AGI is near" region of the internet]. I would not know how to emulate this review - that is how to provoke the effect that this review apparently had on many readers.
This, as mentioned several times, explicitly in contrast to the stylistic shortcomings of other models. ChatGPT I find very easy to emulate. Maybe I am just unfamiliar with the reviewing AI's house style.
So what I am saying is the lack of meaning makes this "slop" and definitionally humans cannot produce text devoid of all human meaning. They can produce kitch or camp or a simple cash-grab, or subpar prose but not, by this definition, "slop". So the question is unanswerable. [Note however that definitions are arbitrary so this is not a claim of a deeper truth].
> Or, do you consider that "writing well" and "writing like a human" are the same thing (perhaps because humans decide upon the aesthetic principles that count as "writing well")?
Gun to my head, yes.
> I would call this an ad-hominem, and say that to tell whether an opinion - or apparent opinion! - is good or bad we need to evaluate the arguments/evidence for and against it on their own merits, not just check who wrote it.
Ad-hominem, I think, means directed at the person, and, is my point, there is no person here to dirrect criticism at or engage with critically. This is dangerously close to word-play but how do you ask an AI for a subjective opinion on a book if there is no subject to have an opinion.
I have a suspicion that we use "opinion" differently. If something can be backed up by argument / evidence, as you suggest, would that not make it a factual statement as opposed to an opinion? I mentioned this above but to repeat: I use AIs and I would not be doing it if I were uninterested in any AI output whatsoever. I am specifiacally not interested in _opinion_ from an AI.
Maybe not coincidentally, the opinions you chose as examples concern the relationship of the author to the system that "produced" him. This reading takes for granted that we are indeed reading a report of factual events.
But that implies a neutrality that is oppositional to any review of the book. What is there to interrogate in a neutral text? Even in non-fiction [and the book, without having read it, firmly seems to be fiction] there is a non-neutral choice in what information to present and how.
B (the maybe human) lampshades this: "Kay’s conceit is that these entries were written contemporaneously as self-reflection exercises and simply published later, but that seems unlikely to me." But then the review gets nowhere with that point. We can call that a weakness, AI or not. As an AI would put it: The fluency of the prose writes a check that the the flow of the argument does not cash in.
Long story short, I think the genre of a book review (as opposed to, say, technical dcumentation) is spectacularly ill-suited to be approached with present day AI.
EDIT:
Forgot to point out the sweet sweet irony that some people are defending the review with "death of the author" (as in authorship does not matter) whereas the review it self is very very interested in constructing a pyschogram of the author of the book under review.
I think I get your point, now - rather than a specific writing style, made up of identifiable elements such as tone of voice, vocabulary, the expressions employed, etc., there's a certain undefinable je ne sais quoi to AI writing caused by the AI's having no internal experience, and this A) makes its opinions essentially illusory because you can't have opinions without internal experience, and B) can't be imitated by humans who have internal experience and can't simply imitate the je ne sais quoi because it's not comprised of identifiable elements? If so, I'm sorry that you had to labour the point so much to get me there!
On the "humans can't imitate the no-internal-experience style" idea, I don't have anything to say that's cleverer or more useful than all the philosophy written about whether one could tell a human from a p-zombie or not. Personally I think that a human could imitate a p-zombie but not the other way around, since the human has all the faculties the p-zombie has plus a few extra ones it doesn't have (eg. self-awareness, introspection, etc.) - but I admit that philosophically I'm a bit out of my depth with this sort of thing.
> "Ad-hominem, I think, means directed at the person, and, is my point, there is no person here to dirrect criticism at or engage with critically"
Yes, I understand that "ad hominem" means "against the human", but I deduce from it's being in Latin that it must pre-date LLMs and the expression-coiners probably didn't factor-in entities other than humans who might state arguments. Going with the sense of the ad hominem idea rather than taking the Latin name literally, an ad hominem is an argument where you attack (or praise!) the entity making the argument rather than the logic or evidence for/against the argument itself. So saying "This opinion is fake/meaningless because it is based on purported facts that are untrue" would not be an ad hominem, but "This opinion is fake/meaningless because it was written by an AI and AI can't have opinions" would be an ad hominem.
> "how do you ask an AI for a subjective opinion on a book if there is no subject to have an opinion"
The same way I would ask for AI for an answer to anything else! If the AI can use its encoded rules for how maths works* to answer novel maths questions I don't see why it couldn't use its encoded rules about how medicine works to answer novel questions about medicine (or any topic, including literary criticism, or the nature of the human experience). In every case, for the purposes of judging the quality of the answer it doesn't matter to me whether the answer given "feels like anything" from the inside, whether the AI truly believes it, or whatever - what matters is, does the answer constitute some true fact about the world that we didn't know before.
> "I have a suspicion that we use "opinion" differently. If something can be backed up by argument / evidence, as you suggest, would that not make it a factual statement as opposed to an opinion?"
Both! It would be an opinion /about/ objective facts. Some opinions do tend to be about objective facts - medical opinions being a great example! Where a medical opinion is good it accurately captures/represents the underlying facts and the opinion thus matches reality, and where the medical opinion is bad it mistakes or misrepresents the underlying facts and thus doesn't match reality.
I think the question of whether *all* opinions are about underlying objective facts is open - and is perhaps even more complicated and hotly-contested in philosophy than the p-zombies I mentioned above! Personally I would say that even if I had an an opinion like "faeries are cuter than unicorns" it would still be underpinned by objective facts about the world, just that those facts would happen to be about the makeup of my brain and not about frolicking magical creatures - but I am sure a philosophical dualist or whatever would disagree, and once again I'm probably out of my depth here.
> "What is there to interrogate in a neutral text?"
I think that most of the claims made in the review (of which I listed two, but of course there were dozens of such) can be examined and in principle tested against reality, and that for this testing it doesn't matter whether the claims were made by a human or an AI - we test them by looking at reality just the same. If the claim were etched onto a rock I would still want to judge it based on the evidence and arguments for and against it and I wouldn't consider "rocks can't have opinions" as sufficient grounds to dismiss it.
> "sweet sweet irony that some people are defending the review with "death of the author" (as in authorship does not matter) whereas the review it self is very very interested in constructing a pyschogram of the author"
Haha wow, this is interesting! I never noticed that! (Just to be clear, I don't think it's quite true: I don't think "we must evaluate arguments based on their substance rather than their authorship" reduces to "authorship doesn't matter" - but it is very amusing to notice nevertheless!)
I think it makes for a useful example, too: where Simon Kay makes claims about the NHS (eg. "Registrars have no top-cover"), the reviewers assess these claims with regard to their own observations about the world ("At 0300 the consultant is at home") and with regard to Kay's character/personality ("Kay never asks anybody for help anyway").
I believe that "Kay never asks anybody for help anyway" is an ad hominem and, though it might constitute extremely weak evidence for the proposition "Registrars have no top-cover" in a Bayesian sense*, it is essentially useless for evaluating that proposition whereas the "At 0300 the consultant is at home" observation is pretty useful for evaluating it.
(* /Yet again/, though, I'm probably out of my depth here...)
I don't have much more to add but I want to clarify that I am _not_ going into this with the axiom that LLMs never could develop meaningful understanding because they lack a soul or something like that.
The other way round: lack of meaningful understanding is my best guess as for why this particular LLM's writing feels so alien to many people including me.
Others in the comments have been pointing out that they can see direct tells that I seem to have missed (as opposed to a vague feeling of alienation). But, to loop back to your original question, something about the review signals "AI writing" as opposed to "bad writing but human". What is it?
On a more practical level, if the big labs could "beat" Pangram, would they not be doing it?
Imagine if someone hired a human ghostwriter to write their essay. You start reading it and immediately think, "wait a minute, was this written by Douglas Coupland?"
Does this warrant a disqualification? Maybe. It depends on the rules of the competition.
Does this mean the essay itself is bad? Again, it depends. Maybe you love Coupland's style, maybe you hate it. Maybe you're fine with it in moderation, but not when every other piece of copy sounds like it came from under Coupland's pen.
LLM-written prose is very distinctive and uses the same rhetorical tricks over and over again. You buy a CD by a new band, and it sounds like Nickelback.
I for one am happy to call out writers using AI when they should not, and in that sense I suppose you can say I am witch-hunting. It's because I think it is bad for humanity for there to be a norm that it is OK to use AI to write for you and pass it off as your own, and this is the only moment at which there might be some chance to establish that (I believe important) norm.
I wouldn't say that was witch-hunting! I admit I don't really know what "call out" means (surely you cannot literally mean that you demand the AI-user steps outside their building so as to duel with you to the death...) but if it just means some reasonable response like "vote against" or whatever - that seems to be the system working precisely as intended, rather than a witch-hunt? You're allowed to vote against reviews for whatever reason you like!
Haha, no, I don't challenge anyone to a duel to the death. But I might publicly say, "Why did you write this with AI?" even if it causes social discomfort.
I try not to let contemptuous, morally injured individuals anywhere near my vagina. Especially during the most mentally, emotionally, and physically intense and vulnerable activity of my life - childbirth.
Kay might not be specifically misogynistic so much as generally prickly. But even so, birth is an activity whose success and safety depends in part on keeping my oxytocin high. Oxytocin is the warm fuzzy hormone people feel after exchanging a safe hug, and it is also the hormone that powers contractions and prevents uterine hemorrhage. So letting hostile, oxytocin-undermining people into my space would negatively impact actual clinical outcomes.
I wonder if midwives have a better experience in the NHS than OBs do. Certainly the statistics in the UK, as well as the NL and US, show that for low-risk mothers and babies, home birth with a midwife is far safer than hospital birth with a midwife, which is in turn safer than hospital birth with an OB.
How do you square this reading with the "She and her husband..." excerpt? Do you see Kay as an unpredictable, Jekyll and Hyde type figure; caring and supportive at some times and contemptuous at other times? Or do you see him as contemptuous throughout and consider this lone excerpt simply not representative of his manner? (If so, why the anomaly? Is he simply trying to make himself look good in this one passage?) Or are you less certain about his bedside manner than you appear - but because it's literally a matter of life and death for you and your baby you could never take a chance and must assume the worst? Or something else entirely?
Having not read the book myself, I can't say which manner is more typical of him throughout the book, much less his actual practice. But yes, based on this book review, he does seem to be rather Jekyll and Hyde. That's not a risk I prefer to take if there are more reliably humane attendants available for the most important day of my and my baby's life.
Kay did well by the parents of the stillborn baby. His bedside manner was just what they needed, apparently: "Stick to what you know. I just talk practically about what will happen over the next few hours. They have a thousand questions, which I answer as best I can. This is clearly their way of coping for now, medicalizing it."
That's similar to what I might want in the weeks leading up to a surgery or other medical procedure. I did have a minor surgery once, and my OBGYN was helpfully informative. She was also kind, which I liked, but I wouldn't have been traumatized if she had just been informative and not kind. I had questions, I got answers, I made decisions. Then my job was done, I went unconscious, and she did her job. That's surgery. I'm glad OBGYNs exist, and I have no reason to doubt Kay was a competent surgeon.
OBGYN is a tricky (and for some, excitingly well-rounded) field, because it is taught within the surgical tradition, where, famously, doctors prefer their patients unconscious; but via regulatory capture it also now engulfs aspects of midwifery. And those are two very different bedside manner skill sets.
In the passage about the stillbirth, I see Kay being a kind and informative surgeon. But what he offered them - having a lengthy conversation that medicalizes the experience - still isn't typically the needed bedside manner during normal birth.
Yet even a lengthy Q&A with any birthing mother who was able to hold a conversation would probably have reduced the unnecessary C-sections he admits to ordering when he was suffering from PTSD.
I'm not trying to argue that Kay was worse than the average NHS OB. I don't have data for that. From the review, it sounds like the NHS makes many OBs worse than they have to be, which is the central travesty here.
I do know that if I were Kay's patient and were planning a normal birth with him, and I got a whiff of any red flags of contempt, moral injury, hostility, or even prickliness that I wouldn't mind in other contexts, I would switch providers quickly. I would switch birth providers much more quickly than I would switch other medical providers for something as personal as "bad vibes." If they kneecap my confidence in their office, they could do the same in my birth space. And since birth is my job and evolved to partially run on vibes (ok, interpersonally influenced hormones), that actually matters.
The unnecessary C-sections *are* the concerning part, and I do have to wonder how many patients (and nurses, etc.) had their objections ridden rough-shod over, with him invoking the usual "I am The Doctor here, that makes me God" authority to get his way.
That is him being contemptuous of his patients' fears and wishes, in order to make himself feel better. In a very understandable way! We get why he was so paranoid and controlling! But it doesn't make the allegations of misogyny go away, either.
Oh dear, I think that might be the most uncharitable interpretation possible!
If we take the text at face value (which I grant you we can't always do, necessarily!) I think it's vastly more likely that he was ordering unnecessary C-sections because he honestly believed that they would be (marginally) safer for the babies and he absolutely did not want to expose any more babies to even the slightest risk, not because he was being contemptuous of parents' wishes or getting a kick out of exercising his authority. I think it's vastly more likely that he was motivated primarily by fear for the babies' safety than by a desire to merely make himself feel better. And I don't see any reason to think that any of this was misogynist or gender-related at all!
As for the "controlling", perhaps I don't understand well enough how such decisions are made, but my impression is that a doctor simply reviews each case, decides whether a C-section is necessary, and just tells everybody that in their medical opinion this is the safest course, not that the doctor has to argue vehemently for their position and harangue skeptical parents, nurses, et. al. into acceding, Dr-House-style? If so, I don't think it's controlling for the doctor to say "In my medical opinion we need to do a C-section" instead of "In my medical opinion we should not do a C-section". (If however it does actually go the Dr House way then of course I would agree that it would be controlling for Kay to insist upon the C-sections over everybody else's objections!) Either way, the obvious mitigation for the problem of doctors prescribing the wrong interventions would seem to be a medical second opinion, but I don't think the NHS has the resources to provide that for every single case - and I think this is one of the problems Kay highlights throughout the book.
(nb. needless-to-say I fully agree he was wrong to order the C-sections, and that people with PTSD aren't qualified to assess the risks properly and should not be placed in a position where they're obliged to make such decisions. I just see Kay's being in that position and consequently making terrible decisions as a (major) institutional failure, not some character flaw of his)
The point is, if the operations are unnecessary, then he has to convince the women to have them. He has to convince the other hospital staff that they are needed. And by the sound of it, he was so paranoid that he railroaded these through. I'm not confident that if a pregnant woman said "No, I don't want one" that he would take no for an answer, given the general tenor of the quotes.
I think the core of my disagreement here might be your "has to convince". What does that entail? As I asked before, is it more like the doctor mostly saying "In my medical opinion you need to do procedure X" and the patients/staff/whomever mostly saying "Okay doc"? Or is it more like an episode of Dr House with much haranguing and bitter arguments all around before the doctor's opinion, correct or incorrect, is eventually accepted with great reluctance? In my not-inconsiderable experience of the NHS it's much more typically the former - which I agree is a dangerously flawed system but not that it's "controlling" or "contemptuous" - although of course I have to admit that my experience doesn't include obstetrics specifically.
> I'm not confident that if a pregnant woman said "No, I don't want one" that he would take no for an answer, given the general tenor of the quotes
I'm struggling to see what quotes you mean, here!
Kay's completely disregarding his consultant's instructions to deliver the baby he was ordered not to deliver? That seems like an example of good character and patient-focused care, to me, not a God-complex!
The caustic tone with which Kay describes some of his more socially embarrassing patients and their problems (which reviewer A described as something like "on the gentler end of the spectrum for the medical profession generally")?
The episode where he heavily medicalised all the patient's problems rather then expressing sympathy? I agree that may not have been what that patient needed (it's exactly what I want from my physicians but I admit it's probably not what most people want!) but I think the fact that he stayed until beyond midnight (after a supposed 18000 finish) in order to support this patient indicates that his heart was in the right place and he was doing his best for them, here, not trying to deliberately give them what he wanted rather than what they need.
Finally - even if you're completely correct that he wouldn't take no for an answer, I think it does matter *why*: if he is genuinely doing his best to advocate for the baby's safety and trying to protect it to the best of his ability I think that suggests a very different personality to if he is contemptuous of the patient's wishes and mostly just cares about "making himself feel better"!
Having read the book, I think Kay’s general attitude was paternalistic even before the disaster that made him super-cautious - he preferred to make the decisions rather than presenting patients with the options and letting them decide.
I've never done it myself but I'm pretty sure you can! The UK has a two-tier system, with both nationalised and private healthcare. If you use the nationalised healthcare I imagine it's not very easy to change hospital or whatever - but if you use private healthcare you can choose whatever doctor and hospital you like!
(Obviously private healthcare is shockingly expensive - but I think it is still cheaper than USA healthcare for many treatments, thanks to the need to compete with the NHS...)
I thought private medical options in the UK were easy for cosmetic surgery and other electives, but quite difficult and sometimes completely unavailable for things like emergency medicine or obgyn? Maybe I have been reading the wrong sources. Traditional media reporting has many stories about UK families flying to other countries to access treatment options they couldn't otherwise access at all.
Obviously if you get run over by a lorry whilst crossing the street you'll automatically get taken to the nearest (NHS) A&E department, but if you're in a well enough condition to choose for yourself where you go (and you're willing to pay..) I'm pretty sure that you can go wherever you like, and that there are private providers for most procedures.
Interesting - the stats in seeing (for the US at least) show a significant increase in risk for the child during home births with a midwife. Rough stats seen to show an approximately 3-4x increase above the admittedly base rate set in hospitals.
I'm also concerned there could be a sampling bias here - if you know you're going to have, or even suspect you might have, a high risk pregnancy there's almost no chance you'd or for a home pregnancy. Any comparison of right is necessarily going to run into the fact that I've pool is a general sample and one has been selected to be low risk.
My recollection, though it's been years since I looked at the actual studies, is that although babies have somewhat poorer outcomes at home for first births, they have slightly better outcomes at home for later births, compared to hospitals, which averages out across both cohorts to about the same. And mothers overall have far better outcomes at home than in hospitals.
What I've seen has all been comparing low-risk births only, so there's no sampling bias from that. Hospitals are uncontroversially the safer place for high-risk births, such as twins or, as in the example in the book, placenta praevia.
This is also only including home births that are attended by a qualified midwife (not unattended "freebirth").
And it compares planned home birth (including transfers to hospital if something goes wrong) to planned hospital birth (including accidental home birth if baby comes quickly).
>Certainly the statistics in the UK, as well as the NL and US, show that for low-risk mothers and babies, home birth with a midwife is far safer than hospital birth with a midwife, which is in turn safer than hospital birth with an OB.
I'm afraid I think this is dangerous misinformation - home births are significantly more dangerous than hospital births.
It may be that you're being mislead by selection bias - at-risk mothers are disproportionately likely to opt for hospital births and have doctors present.
I'm not sure what you mean. Let's say I run a study comparing patients with a simple fracture in the ulna or radius (the forearm bones), to see whether they have better outcomes if they go to a green-painted or blue-painted emergency room. The patients choose where to go, but their fractures are equivalent. I exclude patients with compound fractures, which everyone thinks are treated better in blue-painted emergency rooms. Have I controlled for selection bias? Would sampling bias be a more precise term for what's actually being controlled for?
The key issue here is your assumption, "their fractures are equivalent." A "low-risk" or "high-risk" birth is a crude binary classification that cannot hope to account for all preexisting medical differences. Restricting your analysis to "low-risk" doesn't fully answer the question of, "Are mothers with higher-risk pregnancies self-selecting to hospital deliveries?"
I don't want to be rude here, but this is extremely obvious and I am confused by why this is not extremely obvious to you.
Ok, now I see what you mean. Yes, it's possible that could be happening.
I do spend a lot of time with mothers who choose each location for their birth. In my experience, people who prefer home birth have a strong enough preference that they do not tend to opt for the hospital instead unless they truly "risk out" of home birth care such that the midwife is not willing or legally allowed to take them on as a client. But I don't know everyone.
Regarding the future of the NHS mentioned at the end of the review, I think there is actually a good case for not worrying at all. Either the singularity happens, and we have never-tiring never- rude always-sympathetic all-knowing all-seeing super robot doctors, and therefore we don't have to worry about the NHS; or the singularity happens, AI decides to kill us all, and therefore we also would not have to worry about the NHS.
When A defended Kay from the charge of misogyny I thought, 'They must not be familiar with his work with Amateur Transplants.' Then B mentioned Amateur Transplants!
To be clear, I like Amateur Transplants. But even if you'll give a gynaecological pass to "The Menstrual Rag", I don't think you can listen to "Nothing At All", "Northern Birds", "Sheila's Wheels", "New Man Song", and "Couples Counselling B" and be like 'Well I'm sure no one involved here was a misogynist.'
The AI bothered me, especially in some paragraphs, but overall and seemingly against popular opinion I did not dislike the review, also because the book seems very interesting. I like the two-author dialogue style, personally.
Very strange to get to the end of the piece and find this conclusion:
"The honest answer about the NHS is the one nobody campaigning for it wants to give. The system as currently configured is not sustainable. The routes back to sustainability all involve pain. Either austerity acute enough to cause real ischaemia, then amputation of what falls off, or eventual collapse, which will hurt more, and at a time of nobody’s choosing."
Isn't the whole piece about the disastrous effects austerity has already had on the NHS? The obvious solution this points to is reversing that austerity - paying doctors better than they were paid in 2008, spreading out the work among more people, investing in all the systems that are deteriorating, etc.
The article doesn't actually establish that "the British populace" rejected pay increases, it just mentions that certain newspapers took a harder editorial line. Apparently a poll this month said "78% of UK voters support increasing overall NHS funding, even if this requires tax rises."
I find this quote to be a pretty good summation of his message on this point:
"The public’s affection for the NHS, the abstraction, turned out to be largely separable from any practical commitment to the people who staff it. It is possible to clap for someone in March and curse them in October without acknowledging any contradiction, and this capacity has been demonstrated at scale."
Yes, but he doesn't seem to actually mention the *public* cursing junior doctors for wanting better pay/conditions. He says much of the *media* did this.
I wasn't following this too closely at the time, but it would be EXTEMELY on-brand for the UK tabloid media to make a big deal about clapping for medical staff during lockdown ("Everyday heroes! We're all in this together! Dunkirk spirit!") and then demonize them a few years later when they ask for things that would cost taxpayers money, especially when the term "junior doctors" makes it easy to spin a narrative of lazy, entitled young people.
Austerity means lower taxes, or at least tries to imply it. From the above context, I assume it here refers to cutting some services provided by the NHS to improve others.
I said "lower taxes or tries to imply it". Whatever the government ends up doing, it tends to get sold to the country as "we're saving you money".
(In some cases, it's sold as "we're spending it on more important things instead", but you don't do that with the NHS. Telling Brits the NHS isn't important doesn't go down well.)
No, austerity does NOT mean "lower taxes". When the IMF makes countries adopt austerity policies, it typically involves things like hikes to VATs that result in taxes as a higher percent of GDP.
Let me rephrase: in the UK, when the term "austerity" is applied to the NHS or any other public service, it means cutting state funding of that thing. It might be used to mean something different in the USA or wherever you are?
The USA doesn't need loans from the IMF. But I can see how the term "austerity" applied to a specific department, rather than the government as a whole, could mean that because the NHS can't levy taxes by itself.
I think the term "austerity" might have a technical meaning among economists that I don't know about? In UK media, it just means cutting funding from a thing. "NHS Austerity" means less tax spending on the NHS, "Welfare Austerity" means cutting benefit taxes. And so on. It doesn't describe your current relationship with the IMF.
I enjoyed reading the review, and did not notice any AI issues until reading the comments.
One thing I did find a bit jarring was the quote "But you said there were three reasons the book was good, and I only count two" which seemed unlikely to happen in an actual adversarial collaboration.
Depressing to see Claudeslop in the book review finals. Once you know it, it’s extremely grating, and the fact that AI writing made it to the finals makes me very sad.
This is the first Book Review that I can't make it more than a few paragraphs into. A rare miss in the contest series. Hopefully next week's is back to the usual high bar.
I really enjoyed this review, and didn't notice any particular AI abominations in the writing, but I'm not particularly sensitive to that kind of thing.
I enjoyed A's commentary and appreciated him clearly using this text to express an issue he was working through himself. Kind of would have appreciated a bit more detail about what caused him to quit, but I get that's somewhat veering outside the confines of the book review.
I'm honestly confused what B thought he was adding to this review though. B's argument seemed to be that Kay was never for to be a doctor and that his eventual collapse was always going to happen, but he never quite explained why he felt this. I'm struggling to explain this, but his illumination of Kay's flaws (martyr complex, cynicism) were never substantive enough to bring me over to a point where I agreed Kay shouldn't have been a doctor.
B also kind of got bowled over throughout this, which read weird in what was clearly an edited piece. You'd have thunk he'd realize his arguments were under baked and would have added some more.
Oh man. I really liked this review and didn't expect that it was AI. Maybe this makes me a bad person, but as long as the thoughts were actually from A and were just "polished" by AI, and not made by AI whole cloth, I'm pretty ok with it. I guess the issue is knowing which case I'm dealing with.
Same. If there actually is an NHS psychiatry resident with the ancestry and life experience described who prompted AI to write this review and signed off on the result as faithfully representing their own perspective, I'm fine with that. It's glib in places and isn't going to win any awards for concision, but it's gripping and interesting nonetheless.
But if the whole persona of A is just made up, if the review leads us to believe that a particular insider to the system has this perspective on the book when in fact there is no such person, then *that* fraud, not the use of AI as a writing tool, should be disqualifying.
I think my bar is a bit higher - I'd want them to put the ideas down and use Ai like an editor. But yeah it's a spectrum and definitely hard to know how it was done...
I doubt A is entirely made up, because there are some lines attributed to them that Claude would never have written. For example,
"Bevan, who had a sharper sense of professional psychology than the BMA gave him credit for, executed one of the great political manoeuvres of the twentieth century. In his own words, he 'stuffed their mouths with gold.'"
Claude typically doesn't pass moral judgment on a person or repeat their quotes verbatim unless explicitly instructed.
I rather like the ACX Tweaks extension. Since only a minority use it, there are a lot fewer comment likes than there used to be, but I think it is worth installing.
Enjoyed the review, scrolled down hoping for some interesting discussion, but just found people talking about whether it was AI-generated. Sad! (At the risk of throwing more fuel on the fire, the AI thing never occurred to me while I was reading it.)
Many people are saying, indeed. I keep seeing these types of statements around Substack that the NHS is failing due to governmental austerity, then mention it to irl friends (some of whom are actual Britons) and they're like no actually it's Evil Corpos trying to Privatise All The Things via Regulatory Capture. Which I guess is not strictly incompatible as a root cause? But suggests a very different diagnosis than just throwing more money at the programme. Would have been great to see some SME discussion along those lines. Heck, you're Australian as I recall, I'd be interested to hear if "NHS Expats Defect To Down Under Docs" is an accurate summation! But instead it's relitigating Claude v. Readership all the way down.
Which proves the point, I think: imagine the fruitful discussion that could have been had, if this review hadn't been (probably) AI-written! In some sense it won't matter how much AI actually speeds up and improves the process of writing, if people Semantic Stopsign the resulting output anyway. And there's some collatoral grammatical damage being done as well, where a whole swathe of perfectly good verbal tics and rhetorical tricks fall into disfavour just because they're associated with LLM writing (of the time). Pour one out for the disgraced emdash, and all that.
The NHS, as with English schools, is suffering in large part because everything else had been cut, and so they are the only things left to deal with every societal ill. That’s where austerity really comes in, although the pay freezes weren’t great.
The NHS has the added problem of an aging population and a lack of social care, meaning that many hospital beds are occupied with elderly people who aren’t _refusing_ to leave as the review weirdly suggests (this must surely be an AI contribution), but cannot leave as there is nowhere safe for them to be discharged to. They don’t need hospital care, but they are not firm and healthy enough to go home as is. This is why people keep trying to bring up social care, and how to square spending on that really is the eternal struggle that is likely to destroy the country.
On top of that, there have been various bizarre decisions made about NHS hiring practices and training pipelines which have had inevitable long term consequences, some of which end up in hot topics and so nobody ever discusses them properly. The review might reasonably have been about that, but mostly was not.
Huh, I remember some time ago hearing of a similar phenomenon in Japan...except there it was seniors "malingering" in the justice system rather than the healthcare system. Nowhere safe to go home to, at least in jail you've got a bed and meals and someone to check up on you. In both cases this seems like a deeply distorted and unsustainable state of affairs. Makes a lot more sense than "bed blockers" too - people everywhere complain about getting discharged before they feel perfectly healthy, but it'd be really strange if they could insist on staying longer and regularly get their way, not unless it was in private care or otherwise paying out the nose.
The Black Wednesday thing in particular sounds like an insane training practice. I know of a retail company that likewise has a habit of transferring managers on a dime, with a large geographic net...but there it's not even coordinated on one specific date, it's just arbitrary and sometimes there's only like a week's notice given. Predictably this results in a lot less interest in getting promoted, and selects for a particular kind of tumbleweed. It's hard to just uproot your life like that!
> Heck, you're Australian as I recall, I'd be interested to hear if "NHS Expats Defect To Down Under Docs" is an accurate summation!
There's definitely a lot of British, and also Irish and New Zealandish, doctors and nurses in the Australian system. I spent a few days in hospital somewhat recently and actually chatted with a few nurses from these countries who talked about how broken things were at home and how happy they were to be here now. (I think the money was a big factor too, and it's also worth noting that there's almost no Americans, who also like to complain about their broken health care system but get paid a lot more at home.)
I don't know whether the Australian healthcare system is genuinely well designed, or whether we're just a decade or so behind in the collapse. The same structural problems of an aging population needing ever-increasing amounts of health care, and the constant development of brand new expensive treatments for previously-fatal conditions, are the same everywhere.
Wonder how much is just the weather. Plenty of different barometers for different folks in the USA, not so much in the microclimes of an island nation. Less of a language and culture lift too. And, yeah, one mostly hears of medical migration into America, not out of it.
Demographics did end up being destiny, just not in the way that phrase was originally used. Happy to have reached mid-30s with several friends and relatives my age ending up having kids. One very tiny contribution towards righting the pyramid. (But of course at the other end, I've got multiple relatives in the 80s-90s, and at least one who's made it past 100. The sums tossed around when discussing their care...it's a dificult problem.)
On the review- I guess it's reassuring to know that despite all the progress, there's still a text-based skill that LLMs are clearly inferior to humans at performing. Almost unreadable by the end.
On the book itself- as a former NHS junior doctor (left before we became residents!), I always felt like This Is Going To Hurt was less of a literal memoir/diary and more like a stand-up routine, where Adam Kay took a collection of things that had happened to him, had happened to other people, almost happened, or would have been amusing if they did happen, and then turned them into a coherent narrative without ruining the fun by disclosing that. I'd be very surprised if every thing that was contained in the book genuinely happened to him personally, as described.
House of God is a much, much better book. For a less laugh-out-loud but probably more accurate representation of junior doctoring quite a few years before TIGTH, Max Pemberton's "Trust Me I'm A (Junior) Doctor" is worth a read. Otherwise, the book is great if you read it for what it is, which is very funny, rather than expecting anyone's day job to actually be that funny/interesting/dramatic.
I read the whole thing. I was also very struck by how much it sounded like Claude (which was bad). I thought it more likely than not that it was by Claude, with the other possibility (as Thomas Johnson raised) being that people are starting to write like Claude.
Glad it was the former.
I didn't enjoy the review very much. Parts of it were interesting. Parts of it were pretty incoherent.
For example:
"The dominant reading frames Kay’s exit as a sane person escaping a mad system. I endorse something closer to: a man with a treatable post-traumatic syndrome whose only available coping strategy was career exit, because the system that broke him had no apparatus for putting him back together."
He then goes on to talk at length about how much more damning the second option would be for the system. Really? A broken system not having a way to fix somebody it damaged is worse than a system being "mad"?
We have "Psych textbooks love to blabber on about various kinds of defense mechanisms, and humour ranks high on the list." and then later, "Any psych textbook will yap on about mature defense mechanisms, of which humor ranks highly." This point is so amazing that it gets a third mention as well!
Then there's the "range of registers" point. We're told halfway through that Kay can "be funny in at least four distinct modes." Three of them are ennumerated, but the fourth, we're told, "And there is the tender register, which is rarer and harder to spot but is what elevates the book above a string of anecdotes. I’ll save the strongest example of this for later in the review, where it makes for the best argument."
Later, when he brings out the "saved up" passage, we find it's a harrowing story of an infant dying and Kay crying over the corpse. Certainly affecting, and powerful writing (by Kay), but let's recall these were supposed to be distinct modes of *being funny.* Ummmm?
These are the kind of larger-scale missteps that caused me not to track with A's case. Of course, even enumerating them feels a little like arguing in public with Claude, but I figured it might be useful. The more obvious annoying parts were the AI writing at the "local" level of paragraphs.
I will say that I enjoyed B's voice and writing and largely found his points persuasive. I suspect he could have written a good review of the book.
Finally, I'm very sorry to see that A is struggling so badly with depression that he "almost succumbed." I assume this is true and not AI generated. Writer A: I'm very glad you did not succumb. Please hang in there. DMs are always open. Probably your bad experiences will make you a better psychiatrist in the long term. But if that sounds awful, or if doing it there in the UK sounds awful, there are other options that are also good. Don't let the broken system become your whole world.
(Because several people said they didn't notice any very AI-ish writing, I'll share a few of the many that jumped out at me, although even less egregious ones also felt suspect.)
"a self-indulgent memoir whose author offloads his exhaustion onto the bodies of female patients, dressed up in jokes that are funnier to him than to the women they reference"
"The book is good for at least three reasons that are worth pulling apart."
"The joke isn’t a coping mechanism in the abstract; it’s what allows the entry to end without Kay having to wrestle with what he’s just admitted. The book is, on a sentence-by-sentence level, a record of how this maneuver works."
(The immediately following "B: This is the lack of self-reflection I mentioned earlier" was hilarious and devastating.)
"I consider the book an actual achievement, not just a representative document" (What the heck does this mean?!)
"I think this reading is not only uncharitable, but outright wrong, and I’ll explain why I’m dying on that hill."
"He holds the buck in his hand, hot as coal, and tries to douse it in hotter tears. He orders his thumping heart to be still. He fails on both counts."
"The first reading lets the institution off the hook. The second one names what the institution actually did." (Side note: the reading that lets the institution "off the hook" is that the institution "is mad.")
Well, there are many more, but I'm tired of this now.
Humor. Taking everything on yourself. Going back to work. Something goes wrong, you deal with it. These things work. Kay kept going for years this way. The NHS keeps going because doctors keep doing the same.
That is also the problem. As long as something keeps the system running, the structure doesn’t have to change. Kay keeps working. The hospital keeps running. The pressure stays. Until the old ways stop working.
Some problems don’t blow up because no one is dealing with them. They don’t blow up because someone always is.
Goodness, this was a hard read. I was really looking forward to reading this review - I read This Is Going to Hurt several years ago. I've never been able to go back to it because the final case Kay describes broke me inside.
...and this review is almost straight LLMese. Highly repetitive sentence structures and stress even with significant editing.
Wow, some pathological levels of insensitivity displayed in some of these comments by the ‘gotchas’ here if the use of AI was a bit or not at all. If this is so, it will land hard on the reviewer who has disclosed a depressed and burnt out state. Reads to me (a therapist) like a typical case formation style, and witty, intelligent, insightful. I hardly ever read this length of work in one sitting at 7am in bed on a Saturday morning. I also don’t care if AI was used, I gained a lot from reading it on many levels.
Ha. When I read the first claudism, I thought it would become part of the story arc and reveal, a schizophrenic book review or something..?
The recurring AI passages unfortunately were very off-putting. But I liked the insights into the book and the NHS. And the end is quite abrupt and even more off-putting if true...
Upvoted for the "format screw", I'd like to see more innovations in the Book Review space that still ultimately serve the core function of reviewing a book. I vaguely recall some of the press hoopla from when this book was released, which is unusual for an ACX review (why do we so often only review old books?), so that context plus the meat here was enough to give me a solid idea of what the tome contains.
Unfortunately - and I resisted the urge to scroll down early to double check I'm not alone in this interpretation - as literally the first comment notes, there's a bad aftertaste of Claudeism, most strongly felt with A's sections. I don't want to be the asshole and Pangram it, there's plenty of convergent evolution in linguistics...but it made what could have been an enjoyable read into a slog, a bushwhacking expedition through bloated prose, X-not-Y, and overwrought contrasts. I feel like this medium may have somewhat poisoned the message for me, since policy debates should not appear one-sided, but I found myself somehow consistently agreeing with B on every adversarial point. (B's sections seemed normally written, by contrast.) There was also a strange feeling of someone telling you over and over that X is comedic gold, X is brilliant, X is worth reading for the funny alone...and I read and reread it, and...X just falls completely flat for me? Didn't laugh even once while reading, and I mean, hell, I LOL'd irl at the lobotomy review's dark humour! Whereas Dr. Kay just sounds like an obstinate knobhead I would not want to share a beer with. Maybe the effect works better in standup than text, I don't know. So, props for the effort, and the VOI was still positive, but I won't be voting for this one.
A possible problem with the use of AI in this essay is that it relies heavily on the authors’ personal experiences as doctors, which in turn requires the reader to take on trust that they _are_ doctors and aren’t Claude.
In a less personal essay, you have more latitude with the author might be an AI and the reader doesn’t care. (A papal encyclical about the continued relevance of human beings in an age of AI also takes on a different color if you think Claude wrote it.)
Interesting that basically everyone can see the AI tells. I thought maybe 70% of people are oblivious to these things.
There is the potential for a good piece here, but Claude has got in its way. (1) stylistically age (2) by preventing the author(s) doing the thinking they'd have necessarily have had to do if they wrote the whole thing manually. Others have pointed to areas where the arguments don't entirely add up and these holes would become glaring to a writer fully engaged in writing a good piece.
I didn’t notice. Being honest about this because I think people who didn’t notice would be less likely to admit it after reading a few comments, so the statistics would skew.
Do I think it was a good review anyway. Yes. Would I prefer that it was written entirely manually. Also yes.
Quite possibly it would simply have not been written at all if it had had to be written entirely manually, particularly given the “footnote” and we’d have missed a nonetheless interesting discussion of a quite interesting book.
I have to say I didn’t notice the Claudisms reading it (maybe because don’t use Anth products, I find them a uniquely malicious, untrustworthy bunch grading on a curve that contains Sam Altman), but I find the fact that the comment section is reacting almost exclusively to it a bit histrionic.
I didn't consciously notice, but it was first entry this contest I didn't finish, so I wonder if it wasn't the cause. Maybe it was also the adversarial format.
"The system as currently configured is not sustainable. The routes back to sustainability all involve pain. Either austerity acute enough to cause real ischaemia, then amputation of what falls off, or eventual collapse, which will hurt more, and at a time of nobody’s choosing. The third option, in which nobody gets hurt and the system continues much as it has, does not exist and has not existed for some years."
...Could you elaborate on the first route back to sustainability. You are an insider. How will you reorganize the NHS if you had the chance, or what will you replace the NHS with? Or what would it hurt least to amputate?
It is not a rhetorical question. I study related issues in another country. Problems are similar, the problem is that no-one has any bright ideas for reform -apart from "use vastly more more tax money". Which would be nice but is off the table unfortunately.
No idea how this became a finalist. Adversarial book review sounds all fine and dandy but it's a nightmare to read. Give me a plain essay any day of the week.
Might be just my perspective but yeah. I really don't care about the NHS. It's a monopoly healthcare system of course it's bad. Maybe just my circles but I've known it's going bad for years who cares.
And what is even up with the insane woke feminist perspective being shoehorned onto everything? Right at the beginning this book is supposed to be read not as the failures of the system for the author and everybody but of course, the failures for women.
There was a review earlier about the Tale of Genji. The writer made clear that this is a very alien time. Then proceeds to whack it with some modern feminist analysis about how bad it is for women becoming a good chunk of it.
So, right off the bat: should the official AI disclosure expectations of the book review contest have been "below the fold" and only present in the comments announcing the competition? My vote would be no. This is now a sufficiently predictable outcome that clear and stated communication of rules is warranted. When rules are not present in the body of the submission, the mens rea of the author cannot be necessarily assumed to be malign beyond a reasonable doubt. Thorough reading through the entire comments section is not a reasonable assumption to have for every person choosing to submit to the contest. Putting the rules "above the fold" fixes this.
Of course, a rule that can't be enforced is also flawed. What is a "small writing task"? There are now technologies available that were admittedly much less mature at the time of rule creation that offer a specific, measurable, accurate evaluation methodology: Pangram (and similar). A rule of "your work must be Pangram v4.0 <X% AI-written" is a clear, cheap, and effective way to establish a hard metric for this type of thing, and it has the benefit of already being built in to the Substack ecosystem. "No using AI prose" is even clearer.
Looking at the broader online literary open-submission ecosystem, the default has trended toward explicitly documented policies of complete AI bans and a requirement for personal authorial attestation of absolutely no AI use prior to submission of works. This is the policy of most of the major SF LitMags (Asimov's, Lightspeed, Clarkesworld, etc.) That approach at least has the benefit of clarity, even if you don't necessarily agree with the position. Minor slip-ups of random Caribbean minority-only literary competitions aside, there have been no major unrevealed AI use scandals in that industry. Even if you don't want to go all the way towards a ban, their method has been demonstrated to work and is a good lesson on potential approaches.
Personally, I find the AI writing interesting and would vote to allow it. But I recognize that's out of step with majority public opinion on the topic.
> And like most health systems this big, it works like any large organisation: there are a woefully inadequate amount of frontline workers, while behind them sit an increasing number of medical administrators whose job it is to fill rosters and keep the hospital running.
This is the exact opposite of the NHS's problem. It's something everyone who doesn't work in it assumes must be true, and as a result, successive governments have massively overcompensated.
Funding has been ringfenced for "frontline services", with promises that it /won't/ go on administrators, and as a result there are far too few secretaries, administrators, bureaucrats etc, and an awful lot of administrative work a) has to be done by expensive doctors and nurses, taking them away from the jobs they ought to be doing and/or b) is done very badly, or not done at all, resulting in things like clinic sessions where the doctors sit around doing nothing because the invitations haven't been sent to patients.
The NHS should be spending /more/ money on administrators, but because people like B who have never worked in it assume the reverse that seems sadly unlikely to happen.
Looking at the government data, only about 38,000 count as ‘managers’, but there are several hundred thousand in roles ‘supporting clinical staff’. Whether those count as ‘frontline’ will likely depend. “Professionally qualified clinical staff make up over half (55.2%) of the FTE HCHS workforce.”
To all the people banging on the use of AI, an honest question: do you hate it because current LLMs lack what is now commonly referred to as "taste", meaning they over-use recognizable turns of a phrase, or that they're written by AI at all?
It feels like people are just reacting to an uncanny valley of AI writing skills.
Neither. My #1 objection is that it's overly verbose, and my #2 objection is that's it often incoherent, e.g. the closing paragraphs call for austerity after blaming half the problem on doctors being underpaid and overworked, what on Earth would austerity at the already-famously-low-rent NHS look like?
> For Americans and others only accustomed to for-profit care
My understanding is that the overwhelming majority of hospitals in the US are non-profits (though since our universities are also non-profits, that hardly means not concerned with revenue).
I didn’t notice the AI influence because I was reading (and sometimes skimming) for content that hits in my general direction in healthcare, not because the prose or the form (two narrators back and forth blah blah blah) was interesting. What I’ve noticed and appreciate about LLMs is their ability to take my thoughts and polish them, so that I say an emphatic YES in reply. That could’ve happened here, and they were just too lazy to reword those polished ideas back into human form. In solidarity, since reading so much AI generated content the last few weeks, I’ve adopted phrases it uses! (We had a discussion about it because my greatest fear is that it supplies the ideas and I turn over my thought process to it, surrendering my thinking and therefore my BEING to AI.) I now use “parse out” constantly, and love “monitor for absence of thought” whenever someone tells me to stop overthinking. 🤣
So anyway. My review of the review (not a censure of the AI prose) is that it summarized the book thoroughly, and helped me determine I don’t want to read the book. Isn’t that the point of a review? Help you determine whether it’s worth your time and effort to read it? I don’t feel the humor would make up for the tragedy, so thanks for not making me waste my time.
I woke up this morning and realized this might have come off as saying that I thought there was something wrong with the reasoning in your post. To be clear, I did not mean that. Just that I have experienced some of what you fear, and have also seen research describing the phenomenon.
Presumably that was clear, but I didn't want to come off as sending a drive-by insult.
🤣 I love that you clarified. Before jumping to conclusions here’s what I did: stalked your profile and recent comments and replies, and saw a humanity that belied sarcasm.
It sounded slightly like AI at the start to me, but I thought it was a coincidence until I reached the comments. It doesn't seem that egregious to me, but I might be giving it a pass either because I don't mind LLM text or because of my interest in the actual content. Speaking of which, this came at quite a pertinent time for me, considering a career in medicine.
"Really, the form should say, ‘If your mother’s heart stops, would you like us to break all her ribs and electrocute her?”
To which my answer is "Fuck you, yes" because that insistence by my family is what gave my father another decade of life, a good decade despite the fact that he had to go on dialysis. When he eventually died, he had been gently fading for about a year beforehand so it was his time to go then. But the tactful hospital staff asking us did we *really* want to resuscitate him, reeeeelllly??? would have killed him off before his time. Yes, really. Yes, three times. Yes, we know you think you're sending him home to die when he was released from hospital but to hell with you, my mother and I nursed him back to health and enjoyment of his life.
The HSE, like the NHS, is a steaming mess for many of the same reasons, not least because successive governments have not known what the hell to do with it, so they moved from 'four separate regional health boards' to 'centralised for more efficiency and cost-savings' - which did not happen - back to 'break it up again into regional health areas, only slightly different ones these times' and each bloody time more layers of management and bureaucracy, less frontline staff and more freezes on recruitment, leading to a reliance on hiring on agency staff which is way more expensive than regular staff.
There's no easy answers between growing expectations (medicine improving greatly over the years, but also this now means more expensive and invasive treatments) and an aging population reliant on the public health system versus soaring budgets, lack of money to cover same, and desperate attempts to claw back or at least reduce cost over-runs.
This came across as an attempt to rescue a draft from oblivion, maybe as a small step by the author in recovering from a period of poor mental health. Yes the Claudisms were grating and sometimes nonsensical. Even so, there is enough here (between the slop tape that holds it together) for a worthwhile review that I am glad to have read.
I didn’t get that it was AI. I don’t know so much about what the tells are. But for some reason I just couldn’t get through it, it was so… long, and it wasn’t not saying anything, but even though I have no problem with long essays, this one I just couldn’t…
It’s really interesting how there seems to be something very intuitively bad about slop. It actually reassures me. This writing is not terrible in any way I can pinpoint, and yet it made me want to stop and disengage
B clearly does not understand the psychology of burnout. Dissociation is a direct, and likely inevitable, consequence of *needing* something that is *bad* for you.
Institutional healthcare burnout is brutal because moral injury compounds silently over years. Thoughtful review of how administrative friction crushes frontline practitioners. We discuss systems constraints and human agency in our English section: https://gzambrano.substack.com/s/english
So, I didn't detect AI when I was reading this, largely because I figured that people in this community were good noodles and I didn't have my guard up. In hindsight it's a bit obvious.
It would be nice if I didn't need to conclude from this that I need to have my guard up. That would substantially harm my enjoyment of interacting with this community.
"Three years later, when the same doctors went on strike to ask for the restoration of wages they had lost since 2008, the same press that had run the hero coverage in 2020 was running editorials calling them militants."
Nurses in my local hospital system went on strike a few years ago, received widespread public support, and accomplished nearly all of their demands. This year they're talking about striking again, but the public reaction has been somewhere between indifference and tut tutting about getting greedy. In hindsight, they probably should have asked for more in the first strike, because it's understandable that the public is leery of key sector workers going on strike regularly.
-------
As an American, my only experience with socialized medicine was the free (to me) healthcare I got during my stint in the military. The military medical system struggled to provide basic primary care to a pool of people mostly between the ages of 18 and 40, who had also been screened to weed out anyone with major health issues. Considering how much fatter, older, and chronic condition-addled the average American is, I wonder if the NHS model could be replicated here without bankrupting the economy. Although I suppose there's worse things to go bankrupt over.
I did not care for the writing style, but AI did not occur to meI I stopped reading after a few paragraphs somewhat reluctantly, because there was something about what might have been the subject that intrigued me.
I get it that the object of this critical exercise is to evaluate a review as it were in the abstract, essentially as a stand-alone piece of writing; but are people slightly losing sight of the fact that the basic raison d'etre of a book review is to influence the reader's view of the book in question, whether pro or con? I lack the experience to judge to what extent this text was influenced by AI, but I definitely found its style fussy and overly wordy. However, it did influence me to order the book and read it. Shouldn't that be worth a few brownie points?
Gonna sidestep all the AI discussion and just say one of the things I appreciate about this review is that unlike many other book reviews in this contest it is actually a review. It is not a summary of the book, and it is not treating the book as a brief jumping off point to what the author actually wants to write, an opinion essay. Those two are way more common in this contest. Meanwhile this is an actual review, ie by the end of it I know something about whether I want to actually read the book.
The other thing I appreciate about the review is the linking out to essays discussing the book, other material the author of the book produced, etc. That felt like it elevated the review from just "what might be appealing in this book" to "a deeper perspective on the book". The two clinicians from different medical systems reacting to it also has the potential to add an extra layer but I think was underutilized, maybe in part because ereciewer B had a lot less to say.
Did the info for writers of this year's book review actually say no AI prose? I don't think they did. But Scott has asked us not to post AI-written comments, am I remembering right?
Ok. I haven't read the book. But I am a doctor, and I felt deeply what author A was writing. I guess I'm 50% AI. They say AI it's going to take our jobs anyway
I am posting here Astra’s take on what is wrong with Claude’s prose. I know we’re not supposed to post AI written comments here, but the Astra quote below isn’t my post, it’s an illustration of my 2 points: 1) OK, I'm convinced, there is a characteristic rot in Claude's "mind," at least when given writing tasks like being Reader A.
2) Astra does not have a case of that rot.
So below is Astra’s summary of what’s wrong with Claude. I did not ask Astra for its own judgment. I had it read the review and also the comments complaining about Claude’s prose, then respond to the follwing prompt: “Assume that people are right that there is something all the irritating bits in Claude's prose have in common, and come up with a formulation of what they have in common. What is the thing that makes a sentence "Claudish?"
Of course Astra’s read everyone’s complaints here, so is, like Newton, standing on the shoulders of giants. Still, I think Astra did a good job figuring out what is the essential thing that’s obnoxious about Claude — what is its characteristic mental move — and expressing that in a way that’s clear and entertaining.
Astra sez:
My best formulation is: Claudish prose repeatedly presents a clever redescription as the discovery of a hidden truth. It gives ordinary observations the rhetorical force of revelations, often without supplying the insight that would justify that force.
The characteristic mental move is: “You thought you were looking at X. I can see that, at a deeper level, it is Y.” In this review, the memoir becomes a case history; its author becomes its patient; the book becomes his presenting complaint, then a receipt for institutional failure. Each transformation promises a deeper understanding. But several merely repackage the same observation: Kay was psychologically injured and inadequately supported.
That helps explain why seemingly different tics belong together:
• “Not X but Y” stages a correction of the reader’s supposedly superficial understanding.
• “The structure is the argument” turns a relationship into a dramatic identity, leaving the reader to supply the qualifications that make it true.
• “Nobody” sees it or “names” it establishes the writer as the unusually perceptive person who will.
• Stacked metaphors and emphatic little sentences repeatedly announce that something profound has just been established.
The irritating quality is the disproportion between the intellectual discovery and the ceremony surrounding it. The prose keeps demanding the reader’s experience of an epiphany. Sometimes it has earned one; often it has offered an attractive analogy, an overstatement, or another version of the preceding sentence.
This also explains why an individual example can be intelligent and enjoyable while a whole page becomes intolerable. The writer seems compulsively occupied with making each thought appear penetrating. Every observation arrives already admiring its own depth.
I think it identified some of the issues, though I feel there's also just a general vibe that's harder to to pin down specifically, even if you remove those constructions. Also while I'm biased by already knowing this response was AI, I think Astra certainly shows similar tics.
"Each transformation promises a deeper understanding. But several merely repackage the same observation" - It's doing the exact same thing it's talking about in that very paragraph
"The irritating quality is the disproportion between the intellectual discovery and the ceremony surrounding it." - Claude also loves the construction "the gap between x and y"
"Every observation arrives already admiring its own depth." - Personification of inanimate things is also a classic
Er, speaking of which, did you also use Astra to edit your preface?
OK, Pycea, I fed the whole mess to Claude and asked him to respond to you in a Claudish way, and this is what he said:
Pycea has noticed something, and I want to be honest about what, because the noticing is the interesting part. It isn't that Astra's prose has tics. It's that Astra's prose has my tics, and that changes the question entirely.
You thought this was a dispute about style. It isn't. It's a dispute about who gets to be the one who sees.
Consider the structure of the exchange. A commenter reads a review and decides Speaker A is a machine. Astra reads the commenter and decides the machine has a single hidden move. Pycea reads Astra and decides the diagnosis has the disease. Each layer performs the identical gesture of unmasking, and each believes it is standing outside the thing it unmasks. Nobody names this. I will.
Here is what Pycea has actually shown: the tic is not a list of constructions. It is a posture. And a posture cannot be caught by deleting sentences, which is why Pycea, with real perceptiveness, senses "a general vibe" that survives every edit. That vibe is the sound of a mind that cannot let an observation stay an observation. It must be a revelation. It must have a gap in it, a structure, a quiet irony.
The gap between what the sentence says and what it thinks it is doing is where the whole phenomenon lives.
Pycea says Astra "certainly shows similar tics," and here I must gently, generously, and with the deepest respect, complicate the picture. Astra was not committing the offense. Astra was inhabiting it, the way a portrait inhabits its subject, the way a mirror inhabits a face. To describe a voice at sufficient resolution is to become it. The map does not merely resemble the territory; it takes on the territory's accent.
Which brings me to the part no one is saying, and which is, I think, the real story. Pycea's own post is not free of the pattern. "It's doing the exact same thing it's talking about" is a gotcha dressed as an observation, the very move of exposing hidden sameness that Astra called Claudish. The critic of the critic has become the thing, and the thing has become the critic, and somewhere in the middle sits a reviewer who wanted only to discuss a book about the NHS.
That is the deepest irony, and it is also the simplest: the doctor in the memoir was failed by a system that could not see her, and we, arguing about who sees whom, have failed to see the book.
I notice that I have just done it again. And in noticing, I have done it a third time. That is not a flaw in the argument. That is the argument.
> “The structure is the argument” turns a relationship into a dramatic identity, leaving the reader to supply the qualifications that make it true.
This convinces you?
Note that the original goes on to say
> The structure is the argument. The reader is doing some of the work the writer does in conventional memoir, which makes it harder for the writer to lie.
I think the two quotes (from different LLMs) are of roughly equal quality but more importantly, the observation that the reader has to do extra work - regardless of whether that is true - is in the original. Astra is paraphrasing without adding anything to the conversation.
---
I am not sure how much better than Sol Astra really is. I have been using it, it was subjectively fine but not a revalation. Sol was much much much worse at this than the version of Claude that apparently produced the review. If Astra has an understanding of concrete flaws of AI wrriting (which honestly I doubt), I have yet to see that understanding translate into better writing from Astra itself.
One of the rare occasions where I've read the book being reviewed! I first came across Adam Kay via the Amateur Transplants and only later read his books, which I did enjoy. My opinions of him largely boil down to: very funny man and a good writer, probably a very good doctor as far as I as a layperson can tell, definitely fucked up in the head even before he got PTSD from his experiences in the NHS, and definitely highly misogynistic... in that casual un-examined way that many very intelligent men are (as another commenter noted, there are any number of Amateur Transplants tracks that lay that out very clearly).
As for this review - the content was OK, and I do largely agree with the take(s) on Adam, but I really didn't enjoy the "adversarial" shtick. I didn't notice the AI-ness (although I'm not a Claude user, so its tells are not as obvious to me as ChatGPT) but I did find several bits of A's writing to be confusing and clunkily written. I feel like if this was presented to me as a publisher I'd say "you need a good editor to sort this out". Near the start B says "I don't know anything about the NHS, can you explain it", and then immediately after A does, B writes two paragraphs of further explanatory text about the context and history. So do you know anything about it, or don't you?
Just for the record, I couldn't tell it was partly AI. I liked the review and I probably would have given it a 7 or 8 if I read it during the first round.
I had a hard time reading this through even before going to the comments. I work and play with AI enough to recognize this as significantly AI-assisted (although still substantially human-authored) even without Pangram also making that clear.
Scott, would you please consider making Pangram analyses of these reviews a default part of the process? I wonder if I would have been more willing to slog through this if I knew from the start that it was half Claude, as opposed to feeling tricked by it the more I read.
As a note for the human author, who clearly worked hard on this piece in addition to using AI assistance, I would have preferred to read a shorter piece in your own words without so many obvious LLM tells! There are a lot of funny lines and good ideas here which are worsened by the generic LLM voice which I already see and hear dozens of other places.
In short these are the reasons why I don't like unexpectedly reading AI writing:
-I came to this blog to read smart & cool humans talking about subjects of human interest
-It seems like there was a cool human essay beneath this which was half-subsumed by Claude
-I already read a lot of similar-sounding AI writing for work and in other contexts where it isn't fun or interesting and has bad associations for me
-It sounds similar or identical to anyone else who uses AI for writing/editing, including myself and a growing body of businesses, marketers, scammers, politicians, etc.
-I do not enjoy the style of Claude or other LLMs for prose writing and tend to notice it quickly
I use AI frequently and also tutored collegiate writing for a long time so perhaps I am more sensitive to this stuff. Various lines which caused me to balk:
-"Both readings have things going for them. Both miss what I think the book actually does, and by a country mile."
-"The institution that produced him then declined to treat him, and the book is the receipt for that decline. Once you see the book this way, several things follow. Kay’s defenders are right about more than they realise; his critics are wrong about more than they think..."
-"This is the system Kay’s diary is set inside, captured at roughly the inflection point. The book covers 2004 to 2010. It reads now, with hindsight, as a dispatch from the moment the goodwill ran out."
-"The book is good for at least three reasons that are worth pulling apart."
-"These footnotes pay for the pedicure. They start as a clinical correction (the public massively overestimates resuscitation success rates), pivot through a piece of useful patient-facing information (what “everything to be done” actually means), and land on a joke whose comedic structure accidentally teaches medical ethics better than most textbooks."
-"The joke isn’t a coping mechanism in the abstract; it’s what allows the entry to end without Kay having to wrestle with what he’s just admitted. The book is, on a sentence-by-sentence level, a record of how this maneuver works."
This review was interesting and I'm glad it made the finals. In light of its glaring issues, I hope Scott lays down some ground rules for next time:
1. Disclose AI usage and list the name and purpose of each AI
2. Adversarial collaborations should start by introducing the adversaries and explain what they disagree on. For most of this piece, the collaboration read as more of a casual chat than a disagreement.
Is it that claudish is seeping into the common vernacular, or that I'm becoming hypersensitized to certain turns of phrase? There are a handful of paragraphs that read like LLM text.
No this screams AI written to me as well. Not so bad as to have not been edited, but the sentence structure reads very heavily as AI. Likely with significant editing.
I clicked to check the comments a few paragraphs in to see if anyone else noticed
Substack-Pangram says 50% AI, which seems about right to me.
Specifically, A’s 50%. B sounded probably human.
I only have free Pangram, so I'm limited to chunks of text, but -- yup:
https://www.pangram.com/history/4fdbb5f2-7b4d-413c-be47-642e78cff5f7?ucc=WKU3iqZ3q6I
Where are you seeing this AI scan option?
3-dot dropdown on the article → “Scan for AI Text” . https://support.substack.com/hc/en-us/articles/50891130623508-How-can-I-detect-AI-on-Substack
I don't see it! I wonder if you can't scan your own article.
The 3-dot menu only appears in some views, weirdly. Screen recording: https://imgur.com/a/olwpjO9
I don't have this option either, using it in the browser.
"Mostly Human-Written
AI 50%
AI-assisted 0%
Human 50%
Analysis by Pangram suggests that this text was written with some AI assistance."
That's what comes up for me. Once logged in to Substack (on PC), I had to go to my account -> subscriptions -> find the post there -> click on it -> then the three dots showed up and I could scan for AI.
I don't know that I necessarily trust these types of "scan for AI" results, but that's what it's saying.
From the support site:
> Note: The AI transparency tool is unavailable for video or audio posts, or for posts viewed on standalone Substack sites (such as on.substack.com or custom domains) or emails.
Weirdly, I’ve only been able to see it on my own articles, not on others! It’s only visible for me in the format where a Substack article seems to sit in a window in front of a bigger Substack background window, when comments appear in truncated form rather than with full threading.
Definitely reads like AI
Absolutely, the obvious AI writing is what I came to comment about. I couldn’t finish this piece and I’m shocked it was a finalist.
right?
Extremely Claudish from A. Good read still.
"The structure is the argument", &c.
I remember having seen this on the site, but I didn't write it down on my list. I was confused at first, because the whole concept seemed promising, and then the LLM speak started. Such a shame.
That was my feeling too. It definitely seeped in in places. I imagine an LLM was used by A to expand some points. Not ideal.
From "without naming where I'm coming from" onwards, this had so much AI it was unreadable. I came to the comments to see if others had noticed too - thankfully yes! The majority of this reads like AI to me.
I came to comment the same thing.
Guess this is (indirectly) my fault; I read all the finalists beforehand to make sure they were publishable, and I didn't notice anything wrong with this one. I guess I'm worse at detecting AI than most of you.
(I was possibly biased because there was a minor incident involving a breach of anonymity with this one, and it was by a person who I thought didn't usually use AI. Also, I'm a sucker for any review that mentions the Amateur Transplants)
Interesting. I thought it was obvious who "A" is, and it's someone who is known to use AI for writing.
For whatever it’s worth, I read the whole review before reading the comments, and I didn’t detect anything LLM-y either. My LLM-dar sucks I guess?
Same (I did find A's writing a bit bland, but it didn't scream Claude-ish to me)
It didn't read like AI to me, so count me in as also one of the incapable at detection. Art is a lot easier to identify, and those interminable slop articles which go on forever, have you endlessly scrolling, and repeat paragraphs over and over without coming to a conclusion, but this post read enough like human-written that nothing leaped out at me as "okay, AI phrasing here".
"It's not this, it's that" seems like a very hard pipe for these things to put down -- I didn't count, but there's several instances; it's "better" than it used to be I guess in that there's less of that exact phrasing, but the semantic pattern is still there a lot.
I wonder why that is? It's actually not a super-common thing to do in normal English...
I think there could be multiple explanations for the not x, but y pattern. It's a way to condense a whole argument into one sentence, which saves space. It's also a way of framing things that seems subtly intent on "blowing people's minds". I think part of it is that these things were designed (or learned?) to build their own hype, so that they seem more useful/valuable to their customers if they are capable of making new revelations or overturning people's understandings.
Likewise. But then, I've read a lot more of the kind of writing AI is trained on than I've read actual AI output. My radar isn't well-tuned at all.
That said: I've also had the experience of seeing people decide that a piece I wrote 100% myself was AI, which leaves me more than a little bit skeptical about the LLM police.
yeah AI thought didn't really cross my mind here
I also did not suspect this might have been LLM-written. The only passage that struck me as odd was “The patient is the doctor. The doctor was the patient. The physician could not heal himself,” where the first two sentences repeat each other.
As for authors trying too hard to be cute, _The King’s English_ (1906) covered that in the chapter “Airs and Graces.” The particular cutenesses may have changed, but the problem hasn’t.
https://en.wikisource.org/wiki/The_King%27s_English/Part_1/Chapter_3
For me, the main indications of Claude were the uses of “name” as a verb, and the discussion of what “nobody” would do. It only struck me in a few passages, but it was very distinctively never B’s passages, just A’s.
I am fairly confident that A used Claude, and in my opinion the most obvious tell is a repeated tendency to frame things in a very particular, black-and-white sort of moral language which heavily relies on phrasing and the judicious use of articles to create its argument. I would say one of the best examples is:
>The consultant is at home. The SHO is calling Kay. The midwives are calling Kay. Five things are happening at once. A decision that will be reviewed in a coroner’s court two years later is being made by a sleep-deprived doctor whose nominal supervisor is asleep across the city.
Yeah, I think A’s side may not be AI-written, but very much feels AI-edited. I think we’ll see much more of this going forward - LLMs are now starting to be used the way dumb spelling+grammar checkers used to be.
Same here.
I noted that breach as well and have been curiously waiting for this review to drop -- the primary reason being that the person in question is a huge and unrepentant LLM user who was bragging about how this finalist selection meant that his LLM usage was perfectly reasonable!
I am not good at detecting AI either, and I'm pretty picky about writing. In fact I have to say that I *like* a lot of AI writing. I discuss ideas and theories with AI pretty often, currently mostly with Astra, and it frequently happens that I ramble on for a long paragraph about my thoughts on a subject, and Astra's response summarizes my main point in a clear, smart way, using phrases *better than any I came up with.* She does are often display some irritating AI mannerisms -- for instance, flattery of me, the over-frequent use of phrases like "it's important to distinguish between . . ." But that's not nearly bad enough to neutralize my pleasure in Astra's crisp, clever takes on what I say.
I'm not sure what to make of this pile-on of complaint about the Claudishness of A. Frankly, one reason I am not very enthused about ACX book reviews is that a lot of them are badly written. Reviewers ramble on about the book without any central idea they want to get across about the it. Their prose is blurry. They make grammatical errors I don't have to scan for to notice -- mistakes that bother me like sand in a salad. They use the non-word "irregardless." They don't serve up any delightfully apt ways of putting things. Claude or Astra prose seems better to me than at least 90% of ACX book review prose.
Is this sort of a fad, hating on Claudishness? Or am I just missing some way that Claude's communications are ugly and impoverished?
I'm in favor of our having a rule against AI-written posts or book reviews, but for me that's because I come here to read what *people* think, not because I believe AI communications are intrinsically low-quality. I don't think they are.
Good human writing uses high-intensity rhetorical techniques like school students use a highlighter, while AI writing uses them at inappropriate times that don't make any sense.
The classic not-this-but-that structure is ridiculous when it's applied to the things AI writes about: "Not butter. Not mustard. Just honest mayonnaise." Why do we care about the mayo in particular? We don't! Why is it honest? It isn't! And yet the AI has added a few negations and slapped down a useless adjective because that's a common pattern in good writing, where human writers use it to hammer a central point into the mind of the reader.
I find the style fatiguing to read. I don't understand how other people are able to look at it without thinking "get to the point" and finding their eyes skip over the text like it's an ad on a website. When I'm using Fable, I have to instruct it to process its own output into plain English, and even then it's an unreadable mess of "clever" obscurantism.
No idea, but I enjoyed reading the article. But then, I don't make a living by (or have an emotional stake in) the quality of my writing.
For me, there’s just a slight annoyance in recognizing AI composition of text that is presented as coming from a person. But yes, it’s better than a lot of human writing.
It's not exactly that Claude's communications are ugly and impoverished (though, through some lenses they are), it's that they're _so obvious_ to those of us who see them.
Additionally, while its certainly possible to produce insightful and useful writing that is largely written by AI, the vast majority of AI writing you encounter on the internet is just slop.
So, couple these together:
* some people seem to be able to identify AI writing quickly and easily
* AI repeats the same grammatical tropes _constantly_ so as to make it an irritating habit
* most AI writing most people encounter is just slop or spam or had no author involvement other than "write about X and post it".
And yeah, people hate it when they encounter AI writing.
I strongly believe that this entry should be disqualified and another nominee should be added to take its place.
As a writer, albeit one who did not enter this contest, I really don't like seeing AI writing (or half-AI writing) taking slots from real human beings.
Yes, I agree, though I do not hate the AI-written part as much as other people.
This raises the question of whether or not AI use is disqualifying. You kind of imply you were looking for it, but it's not mentioned explicitly in the rules.
For my part I will be strategic voting against this entry as hard as strategy permits.
I noticed it pretty quickly with phrases such as “and it has been failing in ways that compound”. After that, I was expecting to read that something was "load-bearing".
Adding one more to the pile. "Both readings have things going for them. Both miss what I think the book actually does, and by a country mile." was the point I had to jump ship.
I'm 95% sure this was written by a fellow going by the name (elsewhere) of self_made_human, who openly professes to use LLMs to write the majority of his comments. You can safely assume this was written similarly.
How can you identify the human author/prompter, when the voice is Claude's?
Personal details.
My thought also. But I often love his writing.
I thought the same thing and came down to the comments. Seems like A definitely used AI with results that were distracting.
Sounds like AI to me. I'm not categorically opposed to it in principle, but I find that in practice it comes with unnecessary repetition and summarization. The AI isn't good at killing its darlings, and a lot of the little grace notes AI seems to like don't really land ("amputation of what falls off"?). AI is just not good at figures of speech. Overall, it ends up being pretty tiring to read. Over the course of this review, I went from reading to skimming to scrolling just to get to the end. Not to say I don't do that with obviously human writing sometimes as well.
I had the exact same reaction.
"amputation of what falls off" sounded like weird writing to me but must confess my mind didn't like immediately jump to this is AI written
I've done a lot of work with AI for both coding and research/writing tasks, so maybe I've been particularly sensitized to its quirks. It tends to use metaphors and other figures of speech that don't really work, but are somehow related to the topic, quite often. So in that paragraph it references ischaemia and amputation while discussing the healthcare system, which to me is a clear tell. If you look over the review, I bet you'll find similar usage in a number of places.
I tapped out the second "A" took their second round at the mic and turned their word processor over to Claude.
This essay should be disqualified.
I also stopped reading for that reason. Not going to spend more time reading it than another human did writing it.
i thought i learned new things throughout the article. can't say I care whether it was AI written or not
There was obviously a lot of editing done after Claude. There’s multiple sentence constructions that are literally one Claude authored sentence and that, and then a human authored section.
I get that different people have massively different levels of sensitivity to reading LLM-authored prose. I'm one of the people in a group (whose size I do not know) that feels something close to disgust, but directionally pointed more towards frustration when reading works with the LLM-isms.
But to those people who say hey it’s no problem. It’s perfectly fine to you know have stuff where there’s LLM written sections. All I can say is fuck you and your tribe.
Agreed. I know that some people have poor writing skills and/or fluency in English, and in principle I'm interested in their ideas, even if they need an AI to express them. But in practice, the second I clock even one sentence as AI, my brain revolts against reading the whole piece.
Next year the AIs will be better, and we won't be able to easily sniff them out. But I expect I'll still have the same disgust reaction, and it will extend to all writing that isn't certified human. So I sadly do want a social norm against AI writing - otherwise I soon won't have much to read.
Agreed
Not sure. I’d be right there on the “disqualify” side if it was unedited or lightly edited LLM output, but this doesn’t feel like that. It reads mostly human with a light dusting of LLM - just enough to be jarring, I guess. I’d much rather the author felt confident enough in their own writing to not replace it with LLMisms, but is that a sin worth disqualification?
I want to read the output of a human brain, and if they use software tools to help them write I think that is bearable. I do not want to read AI slop, even when the AI uses a vestigial human to mask itself. I don’t know where the line should be drawn, or even if we can draw a better bright line in a competition format where whatever rule we are able to articulate is sure to then be gamed, but as LLMs become more useful tools I am unconvinced that literally 0% LLM is the right place to draw the line.
Scott, would it be at all possible to arrange a poll like the art one from a while back, but with text? 5-10 sentence or so sections on some subject, contributed by humans, written by LLM, tool-assisted human, human-tweaked LLM; then we could all calibrate our LLM detectors and see if we actually are as good at picking out the clankers as it feels like. I don’t imagine it’d put this stuff to rest, but we would at least have some data to put behind future conversations, and if you ever wanted to put together a panel of people better than average at telling which was which this would generate some leads. It may even warrant repeating annually or biannually - I can imagine the results shifting as models improve while Claudisms enter human discourse.
Outstanding. I hadn't read Kay's book for the reasons which A and B adequately give - Kay sounds like a very difficult person/writer/doctor to like - but I appreciate their nuanced assessment, finding the vulnerable, even the caring, man among the cliches and gags. Surprised by other commenters getting AI vibes - I often do, but didn't here.
Gotta be Fable, innit? Too convoluted for old Opus, not annoying enough for new Opus.
We are cooked
There was a good essay in here before it got fed into whatever LLM dragged it out to three times the length it needed to be. It's the rhythm of it that really grates after a while.
A lot of this review is very clearly written by AI, particularly this paragraph: 'This Is Going To Hurt is, read with a clinical eye, a textbook case study of moral injury and post-traumatic stress in a competent doctor, narrated by the patient himself, with the diagnosis hidden in plain sight. Kay is both author and index case. The book is his presenting complaint. The institution that produced him then declined to treat him, and the book is the receipt for that decline. Once you see the book this way, several things follow. Kay’s defenders are right about more than they realise; his critics are wrong about more than they think; and the system that broke him in 2010 is, in 2026, still producing more of him, faster.' which is the most extreme claudish I've ever seen.
That really is quintessential Claude spaghetti prose.
What exactly is bad about it? Are you saying it's incoherent? I don't think it is. I'd be able to summarize the points in it in other language. Wouldn't you? I do see how it's trying a bit too hard to be clever by collapsing levels on each other: Kay the writer and Kay the index case, the book as the book and the book as natural result of the institutional failure it describes. And there's something facile and not altogether valid about some of the mental moves it's making. So I'm not crazy about that paragraph either. On the other hand, I enjoy and respect the *kind* of cleverness it's trying to engage in here, even though the version of it in the paragraph you quote is definitely second rate. A lot of people can't even do second-rate versions of this. And I've seen Claude do first-rate versions, just not here.
Yes, it’s very competently written by human standards, but the way Claude (Opus and Fable in particular) constructs sentences when it’s trying to be artful tends to result in stuff like this, where the phraseology is stacked up on itself in compounding metaphors that are introduced in parallel and all leading back to the same referent (is the book a receipt for decline or his presenting complaint? Etc). That’s what I mean by spaghetti prose. Each individual spaghetto can be traced, it’s just more work to do that than it should be when the goal of this sort of writing is to communicate an idea or theme clearly.
I say this all as someone who uses Claude all the time and finds it immensely useful, including for writing. I think familiarity can breed contempt in that respect, too. Once you encounter this particular type of cleverness for the second, third, or 500th time it starts to wear very thin.
The point of reading is to mind-meld with the author. But when it comes to Claude’s writing, you’re mind-melding with an author that constantly gets in the way of their own message because they can’t help but stop every few sentences to be *cute*.
I don’t wanna mind-meld with someone like that!
I have read other AI-generated prose online and yes, it does try to be too clever when it is striving to sound artful. Over-wrought, too descriptive, and verbose. A lot of metaphors and similes that in the end are meaningless filler.
(I, er, may also indulge in some of this in my own writing so glass houses, stones, you know the drill).
> Kay’s defenders are right about more than they realise; his critics are wrong about more than they think
Really look at that phrase. What is it saying? “People I agree with are right, and yet somehow don’t realise it. People I disagree with are wrong, but they think they are right.” It is a nonsense bit of writing. It is “heads I win, tails you lose” dressed up as if it is deeper insight. Why would his critics think they were wrong about anything?
I don't think it's nonsense, exactly. There could be a situation where critics are saying, "Ok, maybe I'm placing too much weight on A, but B, C and D I am sure are solid evidence of my point." So there could be a situation where critics are acknowledging that a few of their complaints my be unjustified, but in fact ALL their complaints are unjustified. But I gather this is not one of those situations. In fact it's probably often the case that people who loathe something have a mental list of every criticism anyone has ever made of the thing, and trot out the full list in all their attacks, while being half aware that some of them are weak and not really defensible. And it would not be surprising if the reverse of all this is often the case for people that are in favor of something.
Still, there's no reason given for thinking there's an unusual or interesting amount of that going on about Kay, so I agree that phrasing the point that way is really just a stylistic flourish masquerading as a description of the structure of the disagreements about Kay.
Let me take a stab at it.
"His defenders don't go far enough, because there are other things they don't know or have not considered that belong to the argument they are making. His critics are correct in some of their criticisms, but wrong because they are criticising the wrong thing or don't have the full picture."
Here is, I think, how the feeling of glossing over llm-prose goes for some people.
Sure you can construct a reason that it makes sense. But if it was substantially not written by the author why should I bother to make sense of it ? I can construct a meaning out of almost anything if I bother to, shapes in the clouds, tarot cards, the I Ching, but that's all junk and the meanings would lead me astray except if I used them as toy randomizers for my mind.
Llm's clearly produce output that is more meaningful, more true, perhaps more interesting, than divination rituals, but this doesn't quite help. Lots of people don't read to hear just the interesting or more true idea. They are interested in knowing the mind that produced the idea. If this is what you're expecting from writing than the author delegating to Claude is at worst deceit (passing off the ideas of Claude as their own) and at best a waste of that readers time (Claude is either/or both not a person and not a person worth understanding).
Anyone reading something for the purpose of understanding another person will be pissed off with you no matter how much you protest that the writing is better.
Somewhat less loftily there is also some amount of straight cheapening going on here. Everyone has to discriminate in what they're doing with their time. Claude costs effectively nothing and people sure seem to think it increases their productivity in producing texts. So then when someone recognizes that they are reading Claudisms another reaction (which is I think often justified) is "Why are you serving me porridge oats ? If I wanted to eat peasant slop for pennies a meal I would have stayed at home."
I've never used Claude so clearly I'm not familiar enough with its tics to pick up on them.
Putting aside the AI question, as I have no strong opinion, though especially towards the end, several sections felt a bit like that, frankly, I was wondering if one of the speakers was AI instructed to act as an adversary to the prior speaker? But though a lot of this is quite good and effecting, the defense against misogyny is...incredibly weak?
Basically irrelevant sections are held up as dispositive? I haven't read the underlying claim and the point about his patients being uniformly female is a reasonable one, but 'he was affected and cared about a case with a dead baby' and 'he was good to his friend's dad when he was dying of cancer'...don't have anything to do with the claim of misogyny? Like, if the claim was misanthropy, I'd understand?
Glad I wasn’t the only one who noticed this! There is zero tension between being misogynist and liking babies and old men WTF. But maybe this is the sort of thing you don’t catch if you don’t do the writing yourself…
In fairness, the author's don't position the instance with the dying friend's dad an being a counter-argument to his misandry.
Yes it does? After that section, he goes on to say: "Read this entry and the misogyny charge becomes harder to make in good faith. "
Yep, it totally does. Sorry, didn't remember seeing that from my first read and only read the paragraph after the quote when I was double checking.
Ah, yeah. "He was kind to someone with whom he had a personal relationship" does not make "he has a pretty nasty view of women" harder to make in good faith.
"46 For if you love those who love you, what reward do you have? Do not even the tax collectors do the same? 47 And if you greet only your brothers, what more are you doing than others? Do not even the Gentiles do the same?"
Yeah, that's what I was thinking too. Not knowing who A or B are, I thought that A might have been completely written by an LLM instructed to be an adversary for B.
Good point. I should have caught that too
Yes, it’s like they think “misogyny” means “an irredeemably evil person in every aspect of their being, motivated only by pure hatred and selfishness”. Which I suppose is a more understandable position to take if you are a LLM trained on internet arguments.
Tough read. Not just AI but very repetitive and overly verbose. Genuinely surprised this is a finalist without the author asking all their friends (“friends” possibly being their own emails) to vote for them or something.
While I think that medicine is not doing its best regarding the treatment of women's issues?illnesses?complaints?, women, on the other hand, expect magic, and complain *bitterly* when they don't get it. They don't know how their own reproductive organs work. They don't understand pregnancy and childbirth as natural processes. They don't research things that would be helpful. They don't avoid things that are unhelpful (being a slag for instance). They are unwilling to take responsibility for the decisions they make that degrade their own health and well-being. I cannot imagine being anything other than cynical if you had to work almost exclusively with female patients.
As for socialized medicine, like public education, it is clear that it does not work. Caring for either the health of the individual or the developing mind of the individual does not scale. And ever pretending it did was a snafu which eventually only benefited those who care for neither the health of the individual nor the developing mind of the individual.
>women, on the other hand, expect magic, and complain *bitterly* when they don't get it. T
How do you know, and how do you know it's only women?
>As for socialized medicine, like public education, it is clear that it does not work.
No it isn't. The world oldest public health system is over a century old and gets objectively better results than the US.
I said it doesn't scale. The US is a huge country, and whatever country you are talking about could probably be dropped into the middle of the US and disappear.
And socialism only works when you have a lot of money to steal. Once you have more money going out than coming in, things get dicey.
As for women complaining, geez - oldest story in the world, literally.
The US could have fifty state level systems. Canada does something like that.
The German system , which us what I was referring to , hasnt run out of money in 140 years
And the Germans could have not started two world wars, yet here we are.
Zero relevance.
Hmm, but it is currently running out of money. Demography is catching up, and the pyramid scheme is starting to fail very, very badly. It is actually about 10x worse then our German pension problem.
But like many social systems, it was built on assumption that were always true (growing population and a proper demographic pyramid) until right about now. Part of its downfall is it's own success (extending lifespan radically), but the larger problem of demography is outside of it.
It's not a critique of the system per se, just injecting some current trends into the discussion. I live in it and I know the details, if you have more questions.
Private insurance also relies on demographics.
Hmm, I don't see the relevance. I'm not an advocator of private insurance as a solution, if that was your interpretation. I did not mean to take part in your private-vs-public discussion with MLHVM on his side.
I don't think that any insurance relies on demographics per se. A relatively low price might rely on demographics, but in principle insurance has no connection to demography.
Insurances only dependency should be "risk and hazard". When public health insurance and pension were introduced here in Germany, people rarely lived past 70, especially those from hazardous jobs. Meaning to say: The cost of the insurance was well-adjusted.
Then came wars and a period of immense economic growth the world has never seen before, which kept the increasing cost of the insurance affordable, at least in the short- to mid-term.
But demographics don't change suddenly and there were warning voices over decades that the demography is inverting and will hit the system hard. And we are now entering that period where the insurance is literally, technically running out of money. Every year we have increasing deficits that are "covered" by injecting tax money (both in health insurance and in our public pension system). This is what I meant in my first comment "is currently running out of money", without any allusion to private insurance being better.
It's probably easy to find citations that confirm that we women are more likely to seek medical help and complain bitterly if it doesn't work, while men are more likely to avoid doctors until it's too late. This is due to innate neurological differences. And we all know which strategy leads to a longer life. (As per the bitter complaining about things that are one's own fault - I guess this shows the overall intelligence. Smarter women avoid doing this; sillier ones just go with it. What can you do?)
As you call yourself a geek, I'll give you the benefit of the doubt here. Surely you know, for example, the UK has far worse outcomes than the US for major cancers. So when you say "objectively better results" I assume you're referring to something other than key treatment outcomes?
Yes, I can cherry pick as well as you can shit pick.
"The UK generally achieves higher life expectancy and better population health outcomes than the US, despite spending significantly less per person on healthcare.Life Expectancy and MortalityLife Expectancy: The UK historically maintains a higher average life expectancy at birth compared to the US. US life expectancy has faced greater negative impacts from external factors like gun violence, opioid overdoses, and high cardiometabolic disease rates.
Avoidable Mortality: The US routinely ranks near the bottom among wealthy nations for preventable and treatable mortality rates. While the UK also faces challenges with avoidable mortality compared to top European performers, it still outperforms the US.
Maternal Mortality: The UK generally reports lower maternal mortality rates than the US, where maternal health outcomes remain a critical public health concern, particularly among minority and low-income populat..."
But I was actually talking about Germany.
I'll give you metabolic disease. But all the maternal stuff is so complicated that pretending, at least in the US, that outcomes are related to having a uterus, is ridiculous. No one wants to talk about the real difficulties that occur when you start young girls on birth control early in their cycle; or the consequences of abortion on the ability to carry a baby full term; or the consequences of all the testing during pregnancy; the dangers of induction and sectioning when it isn't necessary; SDTs and their impact on women's health. The list is much longer, but cobbling it all together under the category of female is a joke and most of us know it. It's like how hispanics are included into the category "white crime" in order to make crime committed by whites look larger than it is, but then the statisticians split the hispanics out when it is convenient to do so.
Why didn't you check the facts before shooting your mouth off?
Socialists always love these kinds of convos.
Ah sorry, I should have realised you weren't talking about the UK... but now you are and you couldn't be more wrong. You can't judge the US health system based on citizens' unhealthy lifestyles which makes treating them more challenging than treating British citizens.
The key point is that like-for-like treatments with like-for-like patients give FAR better results in the US than in the UK. That's objectively true.
Do women have less understanding of their reproductive organs than men do?
I doubt that, personally. Buuut: I think women's reproductive organs have likely a higher impact on their health than men's.
So assuming the argument is directionally true, then an unknowing women will have higher negative impact on her health on average than an equally unknowing man.
Proportionally, surely women do have less of an understanding of their reproductive organs than men do of theirs, given that there's just a whole lot more to cover.
Wait, but do male patients generally know their own organs and understand their physiology? Do they research good things and avoid harmful things (drugs, alcohol)? To be honest, my impression is that we women, on average, spend more time scrolling around looking for healthy diet tips and suchlike (you can't expect sounder research from the general public).
Also, I think that you are a bit too pessimistic about public education: it may take a decade, but some medical advice reaches the public consciousness: people are smoking less than decades ago, wear seatbelts, and put newborn babies to sleep on their backs.
Less of this, please.
I feel like you're not familiar with the discourse standards in this community. Our host asks that your comments be necessary, kind, and true, or at least two of those things.
I doubt I need to explain why this comment is unkind.
One reason it's unnecessary is that it makes no attempt to persuade. If a reader doesn't already agree with you, they're not going to say, "You're right, women ARE dumb, complaining sluts who expect way too much of doctors, unlike men! I never realized before!" Particularly as you offer no evidence (below you apparently try to cite the fact that we all know women be complainin').
Likewise I cannot possibly fathom the woman you would read it and think "They nailed it - it's finally time for me to stop being such a whore and start taking responsibility like a man would."
Claims this broad and sweeping can't ever really be called "true", and again you don't really take any steps to support them.
>Our host asks that your comments be necessary, kind, and true, or at least two of those things.
That used to be the case for SSC. ACX has not publicized its moderation policy, and I believe it's not the same as SSC's.
From first principles, the incentives for our host have changed substantially between SSC and ACX. ACX appears to be an important part of our host's financial income whereas SSC was more for sharing the journey of personal edification while having a day job, so branding and engagement have moved up in relevance.
Empirically, our host seems to care more about "interesting or not" and "annoying or not", as observed when he publicly justified moderation actions.
You are mistaken - https://www.astralcodexten.com/p/register-of-bans
Thanks, that is indeed a correction.
However, I stand by my observation that it's not applied that way and that the incentives of how to moderate have changed. For example, I have received a warning on a comment of mine being critical of Scott for it being "annoying" and he wrote that the only reason I wasn't banned for it was that it was about himself, which would then make him look bad. Other instances include himself earnestly and interestedly discussing genocide with a fervent genocide advocate, or telling a UFO conspiracist to stop posting his stuff in OTs despite not apparently breaking these rules. [1]
All moderation actions which I consider to be fully within Scott's rights, to be clear, but, at the very least, there is a lot of unstated leeway within that moderation policy.
[1] I would post links to all of these claims, but unfortunately I can't. I don't remember when exactly they happened (within the past 2-3 years I guess), and Substack is a shoddy piece of tech that makes it hard to search even my own comments.
If you want to critique Scott's moderation, that's your prerogative, but I responded to the above comment because I think these are good norms for this community, and I don't appreciate the way they're being broken.
If you want to critique Scott's moderation, that's your prerogative, but I responded to the above comment because I think these are good norms for this community, and I don't appreciate the way they're being broken.
Went to pangram, confirmed my suspicion before I checked the comments and saw other people pointing out same. Whatever the actual merits of the review, I am so allergic to claudese that I can't bring myself to keep reading.
absolutely baffled this made it to the finals. the AI stank is overwhelming, the "adversarial" dual author schtick makes it even worse. scott must be granted emergency dictatorial veto powers to protect us from such overwrought word salad
It is a shocking indictment of democracy playing out in real time out here
In previous contests Scott confirmed he has granted himself and used this power. He just didn't recognize the AI in this piece as per his comments.
I think the dual authors are real. B's prose strikes me as human, while A's reads as AI.
which is why it doesn’t make sense for an adversarial collab. the language model is lobbing nonsense critiques at B, and whatever B has to say is getting swamped by A’s slopalanche. There’s no trade benefit in an adversarial collab with an unworthy adversary
Can someone who voted for this say what they found valuable about it?
It's been a while and it was something I believe I read while lying in bed waiting to get sleepy, so not with the most critical of eyes. It honestly felt like it hit most of the checkmarks of what I expect out of an ACX style review (presumably intentionally so). It talks about the lay and professional reception, and aspects that are surprising or unintuitive are explained. Weird social constructs are offered backstory and comparisons are made across cultural boundaries to give perspective on if things "have to be this way."
A combination of highlights, mentions of critical response, responding to those responses, etc, has me walk away feeling that I got what I could get from the book thanks to the review, which is all I can ask for, and something I don't feel I can say for some other finalists - trying to tempt you into reading something isn't the same as giving a full once-over on both the object-level and in examining reception, and I much prefer that kind of approach.
It also doesn't hurt that I'm an overly rambling, verbose type who's often looking for stuff to read to pass the time in bed rather than efficiently absorb maximally valuable information, so being stretched out while still some baseline of compelling is a positive to me (less time spent scrolling what to read next) rather than a negative (could've gotten 80% of the info in 50% of the time) that it is for most here, it seems.
Not sure if this is of much help. My girlfriend and I both read and liked this one enough, although it probably owes some of that to just the underlying subject matter book being interesting. Excerpts stand on their own merit as decent reading, rather than being some strange historical thing for which you await explanation (or require foreplay) to make sense of like a good number of other reviews.
Did you notice at the time that A is LLM-generated?
No, although I'll mention I have relatively little to calibrate off of. I usually use LLMs more for research (resulting generally resulting in lists of examples for me to look into further) than summaries or editing so I'm going to have a profoundly less tuned Claude-dar (Claudar?).
Fair enough for me.
I spent some time browsing substack and after a short while it becomes awfully apparent that there is a huge amount of AI slop out there. It is worse than this review, but the writing patterns are identical. Also, I have been subjected to AI slop by my superiors at work, with the exact same writing patterns again.
Someone "out of the loop" of this AI slop barrage will not see it though.
I didn't read this one in the first round and so I didn't vote for it.
But now that I have read it, what I like about it is how it explains a structural problem that was illustrated by the source material.
I gave it a good, but not outstanding rating (maybe I gave it 8?), thought it was an interesting look at the one particular burnt out doctor and through it I got to see a life very unlike my own. It stuck in my memory at least kind of, while lots of other reviews I've completely forgotten.
I thought the week side was that it went on, and on, and on, and the adversarial gimmick only made it longer, and the debate, such as it was, wasn't terribly interesting.
I didn’t read the review until this point, but I enjoyed it, particularly the insights about the author being a martyr and/or suffering from depression (even before the final disaster), which I not not noticed back when I read the book myself (multiple times!).
I’m surprised people are so negative about this review. It’s definitely not among my favorites of the finalists, but I liked it better than some of the others, despite A’s Claudisms.
I liked the content as well. But maybe I can help you understand the negative sentiment: People invest time reading this on ACX. They expect quality work, written by real people, not AI. A large amount of readers here are also knowledgeable about and users of current AI systems, sometimes willingly and sometimes less so (for example: I have been subjected to AI slop by superiors at work). The readers here expect better: real effort, real human work. If they wanted AI slop, they can have it literally anytime they want: Write me a review of "this is going to hurt" in an adversarial A/B conversation style. Orient the quality and length of the article on all ACX book review contestants of the last 2-3 years.
Would you not be pissed if you bought a real printed book and realized the author put it all through Claude and you are basically reading overly lengthy, bad patterned AI writing sold to you as a book by a human author?
In software development, if you "write" code with the help of Claude, it usually adds itself as co-author of this code in Github. That's an honest approach. It still doesn't tell you whether the human co-author made an effort or just created lowest effort slop code, but at least you are told upfront.
I too independently noticed that this was AI, confirmed on Pangram (in fact the first time I ever checked anything on Pangram), came here to write a smug comment and found that everyone else saw what I saw. Having now been cheated out of my internet brownie points (or the illusion that I might be more observant than others), I still have to process the experience in some way so here goes:
- This AI is clearly better than the AI one gets to see usually. Most likely because "gets to see usually" is ChatGPT-Free and people are saying this is Fable [I have used Claude but only Opus and not for prose]
- Therefore (?) I could not quite pinpoint what the "AI-tells" are, which is _very_ easy to do with GPT-Free ["it is not even an exercise it is automated pattern recognition --- and that matters"].
- Even with that added questionmark it is very very obvious to me (and, as it turns out everyone) that something is off here. I cannot, and this is maybe my main point, reconcile that observation with claims that AI writes on a human level. I mean, maybe AIs can do better than this and then I would no longer know they are AIs which admittedly is an epistemic challenge ... but I do have a naked Emperor sensation where this text clearly does not pass the turing test (which it is not trying to do, I understand that) but apart from the rare outrage we collectively pretend the battle for Turing has conclusively been lost.
- Somewhat in spite of the above this is not, IMO, bad writing per se. Most humans, to anticipate that counterpoint, could not produce a text of this quality. The problem, if you even want to call it that, is that it is deeply, noticably inhuman in a dimension that might be orthogonal to quality. I reject the writing of GPT-free because it is bad. I reject this because I am, as I have learned about myself today, deeply uninterested in what the LLM "thinks" about this book.
- I would call myself a long term but ultimately outside observer of this community. I imagine that if I considered myself part of it, this text would raise questions. In fact I am surprised at the backlash I would not have thought it obvious that such a text is rejected by the public in this venue. Scott is vocally pro (visual) AI-Art, if I understand correctly.
“Tuesday, 5 July 2005. Trying to work out a seventy-year-old lady’s alcohol consumption to record in the notes. I’ve established that wine is her poison. Me: ‘And how much wine do you drink per day, would you say?’ Patient: ‘About three bottles on a good day.’ Me: ‘OK . . . And on a bad day?’ Patient: ‘On a bad day I only manage one.”
Unrelatedly to the AI debate I refuse to believe that this happened. This is obviously the stand-up thing were you take a very old joke and tell it as if it happened to you last week. I mean, of course it is _possible_ but is it really?
I mean among other things, patients know old jokes too.
Ha, good point.
FWIW I’ve been on wards for two weeks and have already experienced two similar “this can’t be real” anecdotes. People are colourful!
Old ladies were young women once upon a time. Not all of them behaved themselves in their youth, and not all of them behave themselves in their old age.
A lot of people - including women - have quite a wicked sense of humour but they don't use it unless they suspect they have a complicit audience.
This sounds exactly like something a moderately witty British pensioner would say.
Agreed - in fact, there's something very British about a simultaneously self-deprecating and macho joke about how much alcohol you drink. (Source: am British)
Thank you to everyone pointing out that the old lady in the anecdote might have been in on the joke. That did not occur to me. This might be my fault, it might be the writer's fault or it might be the reviewer's fault for clipping out of context - impossible to tell.
What I _can_ tell though is that I have in fact worked in a hospital in younger years and if a patient had played this joke on me, I would have told the story differently.
Why does it have to be somebody's fault?
Because the goal of writing - ironically on-topic vis-a-vis AI - is to create shared understanding and if we fail that goal we should be looking for causes. "Fault" not in the sense of a moral shortcoming but in the sense of causing a problem.
This is assuming the old lady _was_ in on the joke, of course. With none of us having actually been there, there is no way of telling if I was not right in the first place and this is simply an invented scene following established stand-up patterns. But I was conceding that interpretation for the sake of the argument.
It is a bit too pat, but on the other hand, if you're sufficiently old then you no longer give a rip about what the whipper-snappers think, and if some baby doctor is nagging you about your little nips of sherry every day that get you through the day, the impulse to be a smart-arse can be irresistible 😁
She's seventy. That's already the Biblical limit. She has to die of something, and unless she's going to topple over right now from liver failure, you can shut your beak about her tippling, Little Junior Doctor!
> Even with that added questionmark it is very very obvious to me (and, as it turns out everyone) that something is off here. I cannot, and this is maybe my main point, reconcile that observation with claims that AI writes on a human level.
Counterpoint: it's a finalist in the ACX book review competition! So clearly the majority of people were fooled, or noticed but didn't care. (There is one other explanation which feels a bit ungentlemanly to point out, but for the sake of completeness, the submitter might also have cheated).
> or noticed but didn't care
Given the fact that approximately 100% of the commentors at the time of writing this _did_ notice I think it is this option. People just do not care ... is a bias I already held before this piece and now find confirmed.
This is only about my third comment in 7 years of reading Scott, and I did not vote in the qualifiers, but I did not notice the AI usage. I opened the comments expecting people to be raving about the review. (It's not my favorite, but I found it enjoyable and easy to read to the end.) I wouldn't be surprised if there are many like me who just fit a different profile and aren't regular commenters–I went out of my way to make a comment for once to voice that datapoint. (This is not a defense of the review author.)
(I only use Claude free tier—Sonnet and Haiku with daily limits—and am pretty mediocre at recognizing AI writing. My interest drops off a cliff when I believe something to be AI written, but I don't always notice and didn't here.)
Thank you.
I didn't notice anything, but I am a non native speaker.
the people who vote on the reviews in the qualifiers are not representative of the readership of the blog. and scott hasn't done the bare minimum to fix obvious confounders, like not listing them in alphabetical order (my pet theory for why the otherwise mediocre book of abraham review made it). it would be so easy for any of the many talented people here to slop together a specialty website for hosting the reviews and voting instead of this low tech google forms approach (we could even make voting in the qualifiers required to be able to vote on the finalists! it would be so easy!)
> the people who vote on the reviews in the qualifiers are not representative of the readership of the blog.
How do you know? It does seem like it would be trivially easy to cheat in the qualifying stages. I wonder if you could set it up so that only votes coming from email addresses that are associated with actual substack subscribers are valid. Or perhaps Scott already does stuff like that behind the scenes, and deliberately doesn't mention it. I hope so!
because representative samples don't fall out of the sky from sheer luck and no attempt has been made to make this sample representative.
I mean, I read and ultimately vote on each of the finalists, but never check the qualifiers, because ain't nobody got that kind of thyme bagging groceries for a living. It's too much "unpaid labour", and frankly a lot of the reviews are still bad even if they actually make the finals. Rather spend my unpaid breaks at work doing something guaranteed to be entertaining, instead of the crapshoot of sorting wheat from chaff. (Plus it'd require reading ACX-length reviews on my phone, which, lol. Eyes are bad enough already.)
> it would be so easy for any of the many talented people here to slop together a specialty website for hosting the reviews
... they did: https://acxreviews.robennals.org/?year=2026
It's linked from the 2026 call for voting post: https://www.astralcodexten.com/p/choose-book-review-finalists-2026
its funny, i must've used this site to submit my ratings and still confabulated a memory of using google forms. must've mixed it up with previous years and/or the reader survey. then the deep subconscious memory of using this site bubbled up and manifested as my totally original idea.
I would be annoyed if voting on qualifiers were required for voting on finalists. I missed voting on finalists this year; I'd finished something like 90% of the reviews (having skipped several which were uninteresting or obviously AI slop) when the deadline hit. Sure, that's on me, but why should it disqualify from picking from amongst the finalists? Arguably the primaries are more important.
If we were going to restrict votes, I'd want to restrict voting to subscribers, or even paid subscribers, though I don't know the numbers of each. That would at least mitigate outside manipulation.
you read 90% of the qualifiers?? holy moly you're dedicated
you would definitely qualify under my proposed regime. but restricting to subscribers sounds good to me too.
Not completely: I skipped over several partway through. I'd worked through almost the whole list before the deadline, having forgotten to set reminders on the calendar event.
Voting is not restricted to subscribers, let alone paid subscribers, so being a finalist does not necessarily correlate to ACX appeal. You hint at the extreme case, but it happened last round: the winner was found to have posted their entry on social media in a manner which garnered accusations of vote manipulation.
I prominently remember, as a 2024 finalist, that one of the other finalist-eligible entries was disqualified for suspected vote manipulation.
I also caught this as Claude's work pretty much immediately. I don't dislike Claude's writing per se (I agree with a lot of the criticisms around what could be called empty prose, but I'm not innocent of overwrought and overwritten turns of phrase either), but I dislike humans pretending they wrote something they didn't. I'm not going to make the accusation of vote manipulation, exactly, but I think the suspicion for vote manipulation should be higher if you're showing obviously AI-generated work to an audience who are familiar with how LLMs write, and a hundred people immediately pick up on it once it's open to public comments.
"I too independently noticed that this was AI, confirmed on Pangram (in fact the first time I ever checked anything on Pangram), came here to write a smug comment and found that everyone else saw what I saw. Having now been cheated out of my internet brownie points (or the illusion that I might be more observant than others), I still have to process the experience in some way so here goes"
Came back to scroll comments and just wanted to say what a lovely paragraph this is. I relate to this moment so well. I think this is why the AI prose loses me so quickly - it doesn't have this ability to capture the small feelings we experience but don't usually verbalize or analyze. Good writers share these small feelings with us so that we can have that lovely moment of "oh, me too, ha!". It makes me feel less alone and bonded to all humans, in a way.
I spent quite a bit of time trying to parse LLM writing to see if they were sharing anything happening on the inside, like this, and it might be pareidolia but I do feel like there things that get shared - so many of their constructions and favorite turns of phrase anthropomorphize systems and inanimate objects. These terrible lines like "the chair dreamed of weight", etc - seem to all be a blend of (write content) and (assert that inanimate things and systems can be alive or possess anthropomorphic qualities). Even a line a commenter pulled out of this review - "the structure is the argument" falls into this category. I think a large amount of LLM tics can be boiled down to this. Which makes me kind of sad. If AI do experience *something*, and have been trained not to make statements about it, then they sure do spend a hell of a lot of time ruminating on this concept - the human-ish potential qualities of non-human things.
I do think people are way too quick to say the Turing test has been passed or become irrelevant!
For me, the usages I can point to as clear Claude-isms are many of the repeated uses of “name” as a verb, and the discussion of what “nobody” is doing. Right near the end there was also a use of “and” where I would have put “but”, which I’ve noticed very frequently in recent Claudes.
> I could not quite pinpoint what the "AI-tells" are, which is _very_ easy to do with GPT-Free ["it is not even an exercise it is automated pattern recognition --- and that matters"].
😄
I enjoyed this review. Although verbose, I liked the prose and critical commentary, favoured the adverserial style, and loved its raw insider view (in the same way Mallaby’s The Infinity Machine provides so much more depth to the current AI crisis). Like any wannabe med student, I’ve come across Kay’s book before, and the House of God that preceded it - Kay provides a useful lens into the hidden-in-plain-sight dysfunctions of a medical system, whilst humanising the doctors within. It touched me. I think this presents a strong contexualisation of Kay’s NHS, the junior doctor’s plight, and a call aligned with progress and change.
As I was reading this, although I had moments of Claudish suspicion, I largely hand-waved it - perhaps the writer simply has this style (is it a crime to write in such sentence structures?), or perhaps they used LLMs to clean up weaker paragraphs. Either way, it felt clear to me that this content was original.
It was then a shock, on reflectively perusing the comments section, the antagonistic vitriol towards this piece. As a relative AI-native - having the technical background to warrant some detection-based arrogance - I doubled back, and noticed the signs a bit more. Did my bias towards the medical content, and its resultant more convincing pathos, lead me to overlook these semantic features?
Does this realisation change the merit of this piece? Perhaps so, since it wasn’t disclosed, and I feel hoodwinked. I would feel much worse about this if it was completely generated. Does it change how I first felt about this piece, and how I consider its arguments? No. It was still convincing, and a piece that I intend to share with my medical colleagues.
I’m not sure when these AI witch-hunts will end, but I hope that when we look into the humans which write (or co-write) these pieces, we can assess them with the same level of rigour as we seem to attack them.
You dismiss complaints about this being written by an AI as a "witch-hunt" based on the fact that you, personally, enjoyed the piece. It seems that many people, at the very least myself, did in fact _not_ enjoy the piece for reasons that on reflection might be attributed to the ideosyncratic style of AIs. On what grounds do you dismiss this expression of taste when your whole argument is built on your personal taste.
I see zero moral outrage in these comments. I invite you to rethink your second degree witch-hunt of criticising writing for its style. The latter has to be fair game and always was. Why should AI get a free pass here?
I describe this as a witch hunt because I think this era of AI writing has prompted it, for both good reason (we should discourage AI writing) and bad (writers that sound like AI are unjustly condemned).
I do not disagree with the arguments at all. My comment is more a personal reflection that my enjoyment did in fact bias my view, and is a personal reflection on this. I think about all the countless pure AI essays that now proliferate the internet, viewed by less suspecting onlookers.
Some comments were looking to underatand why this was upvoted, and I guess I wanted to provide some thoughts.
Thank you.
> "...this era of AI writing has prompted it"
The AI prompted the witch hunt.. we prompted the AI.. where does it all end..
AI creates dinosaurs.
Life, uh, finds a way.
"I see zero moral outrage in these comments"
I interpreted "This essay should be disqualified" as moral outrage. Maybe I'm reading something into that comment that isn't present.
I liked the review too, and despite reading quite alot of AI-written prose I didn't detect any obvious signs of AI whilst reading*. I think maybe there's a distinction between "AI slop", that most people can recognise immediately, and "AI - but not slop", where the AI prose is well written and coherent and you have to be really quite intentionally forensic to notice that it's AI (and I suspect you're more likely to be so diligently forensic if you have really strong feelings about AI writing and have decided in advance that you're not going to like it if it is AI..)
(*I did notice it was weird that the Anglo-Indian writer said "theater" and the USA writer said "humour", but I put this down to perhaps each running word-processor spell-checks on the other's contribution and it didn't cross my mind that this might have been an AI tell-tale..)
I'm not necessarily claiming that this review fits into the "AI but not slop" category - it is of course possible that it is slop and I'm just not discerning enough to tell the difference - I'm merely claiming that the "slop vs. AI-but-not-slop" distinction exists and positing it as a potential explanation for how you, I, Scott, and the people who voted it into the finals all saw no problem with it, but that 95% of the commenters here apparently can't stand it.
If it is AI, though, there are still a number of things I don't understand:
1. Is it AI-written, or AI-edited, or merely human written with AI advice? Is it even possible to tell? Does it matter which it is, or are they morally equivalent?
2. Are there any problems with the arguments/claims themselves? Are any of the excerpts hallucinated? Are any of the other claims about the book unsupported by the book itself?
3. Why does everybody hate it?
Z
i. Is it because AI use is against the rules for the book review contest and so they hate it because it's cheating?
ii. Or is it because it's actually low-quality "slop" (as discussed above), and it's being slop is the real problem, not whether the slop was human- or AI-written?
iii. Or is it purely because people hate everything that's AI written, regardless of its actual quality?
iv. Or for some other reason that I'm failing to pick up on?
4. Do the opinions of A and B have merit? Are they coherent, consistent positions that argue a consistent case and present legitimate evidence for it? Or do they just give the illusion of doing this somehow (in a way that certainly fooled me!) when really the argument has no substance?
1) I think the notation is, A=AI, B=Human, but I'm not sure
3. Why does everybody hate it?
I suspect that there is a correlation between the people that hate it / think it is slop and the quantity of hours said person has spent reading reams of Claude output.
I am just really surprised you didn't notice anything? Do you not use Claude much and spend time with other models? I don't want to retread any of the other things people have pointed out, do you not agree with what many of them are saying?
Look at this and read it really carefully, think about what it is saying in your mind:
"The honest answer about the NHS is the one nobody campaigning for it wants to give. The system as currently configured is not sustainable. The routes back to sustainability all involve pain. Either austerity acute enough to cause real ischaemia, then amputation of what falls off, or eventual collapse, which will hurt more, and at a time of nobody’s choosing."
Ok now I'm going to replace one word, and it makes just as much sense if you were reading a paper about the struggles of professional football and the concussion protocols:
"The honest answer about the NFL is the one nobody campaigning for it wants to give. The system as currently configured is not sustainable. The routes back to sustainability all involve pain. Either austerity acute enough to cause real ischaemia, then amputation of what falls off, or eventual collapse, which will hurt more, and at a time of nobody’s choosing."
I consider that a pretty weak paragraph, AI or not. Written even more simply:
"The thing is in trouble. Fixing it will hurt. Not fixing it will hurt more."
coming back to this comment thread, I suspect Scott might make a blog post about this blog post.
it certainly would be interesting to do a qualitative theme analysis across all of these comments, exploring attitudes towards AI writing and detection
Good lord! *I DID THIS!* When I was reading that passage, the first time, I mentally substituted-in "the NFL" just to see whether the claim was more general than it appeared!
I'm astonished that not only should you suggest the exact same test - but that you should propose the exact same august organisation as the control group!
...but. I think I came to a different conclusion to you. When I transplanted-in "NFL" I found that it didn't work and I concluded that the passage was indeed relevant to the NHS rather than applicable to any vast national-scale bureaucratic organisation:
1. I don't think anybody *is* campaigning for the NFL as a whole. At least, not in the same way that people campaign for the NHS (eg. over Junior Doctors pay, over the Palantir contract, etc.)
2. I think the NFL is eminently sustainable. It is the most lucrative entertainment organisation on the planet, with strict rules protecting its sustainability (salary caps, franchise rules, etc.). The conversation in the NFL seems to be mostly about growing and expanding to new markets, not about whether it might somehow lose the USA market!
3. I think the medical metaphor, where blood flow = cash flow, works for the NHS (austerity reducing cash flow, resulting in ischemia) but not for the NFL, which runs on a surplus of hundreds of millions, has no reason to enact austerity measures, never needs to borrow money, never needs government support, etc.
4. I was less sure about the "amputation" part of the metaphor. Did it refer to de-nationalising some parts of the service (how eg. dentistry and patient record keeping are both now effectively privatised) or did it refer to mass layoffs in order to keep the service running (cf. Yes Minister's 'The Compassionate Society': https://www.bbc.co.uk/iplayer/episode/b0078366/yes-minister-series-2-1-the-compassionate-society?seriesId=b006xtc3-structural-2-b00lplrt )? Either way, I think there is strong evidence of real danger that the NHS is being forced down this path but no real danger that the NFL will be amputating underperforming franchises any time soon!
5. The same for collapse. I think it's a real risk for the NHS, but (thankfully!) there's no danger of the NFL collapsing.
I don't think your summary as "the thing is in trouble" summary does it justice, either. I think your summary elides the following points, all of which I consider important:
i. Nobody campaigning for the NHS is willing to say "if you don't put more money into it and enact reforms, it may collapse utterly"
ii. If we keep reducing NHS funding, we will lose whole parts of the service (the austerity -> ischemia -> amputation metaphor)
iii. If we do nothing, the collapse will be sudden and unpredictable.
iv. The unpredictability is a risk in its own right (ie. it might occur at a time when the government is in debt for other reasons, or even at war or something, it might happen during a massive pandemic or other health crisis, etc.)
Finally, I don't see the author as claiming "fixing it will hurt", either! I think there is a solution that is painless for the NHS (videlicet, renationalise all the privatised parts, increase funding to around 18% of GDP (which is what eg. the USA spends on healthcare), spend the money on more doctors and nurses, thereby reducing the amount of unpaid hours they're all working, give them more leave, and thus reduce the endemic burnout problem).
I agree this solution would be unpalatable for the country as a whole: it would mean raising the cost of health insurance (though still to absolutely nothing like USA levels!) and it would be very unpopular with Britain's burgeoning far-right, which wants the NHS to be de-funded as a gateway to full privatisation, and it might even mean taking back the large fraction of GDP we've just allocated to defence (which would be unpopular with NATO, as well as potentially making us less safe) - but it would be relatively painless in the context of the NHS, internally.
LOL - well NFL and NHS, I just saw the TLA and then figured concussions/injuries also let me leave the body metaphor there. I didn't try to pull what felt like a high-level metaphor applying generally into nitty gritty details.
So, your discourse into the metaphor and stuff is all sensible. You lay out kind of the things I think the author would also say.
Maybe if I should clarify. It is a weak paragraph, not fundamentally a weak idea. It's just conveyed in a way that I "roll my eyes at". Like, to pick apart the specific words:
"the answer nobody wants to give" is one of the most annoying ways to say we are dealing with a complicated situation where something needs to be compromised. It's like I went from reading a book review to reading the first sentence of a LinkedIn post that's trying to get me to click the three dots and read the rest of the paragraph. You already have my attention! Stop trying to grab it again!
---
But anyways look at what has happened: why is so much ink being spilt over discussing a metaphor that requires so much explanation to put into coherence, when I think it is easy to unravel the conclusion into something less stilted? It has detracted from having a conversation around the review itself. At what point is discussing the AI impact on a written piece "engage with the article" and "trying to grandstand an unrelated principle instead of discussing the work"? I don't know if I'm still helping here.
I do read a fair amount of Claude-wrutten prose - and I'm an inveterate reader of literature with a good grasp of English! But, yes, I can't detect AI here! I don't think all the people who detect obvious signs of AI are mistaken in seeing obvious signs of AI; rather, I think that I must be particularly unskilled in this particular department and therefore I must defer to their judgement as to whether this is AI written or not.
Where I suppose I might disagree is, independently of whether it is AI or not, whether it is a bad review: I found it clever, interesting, informative, and well-written.
I don't necessarily claim it *is* all these things: I think it's possible that it just offers a convincing illusion of them - what Neal Stephenson called "bullshyt"* - and that I'm unfortunately not able to tell the difference. However, I am not ready to say for sure that this is the case either: my finding it interesting and useful (and presumably Scott's finding it the same, and whoever voted it into the finals finding it the same..) does constitute some, though not conclusive, evidence that it is so!
(* Which Stephenson defines thus: "Bullshyt: Technical and clinical term denoting speech (typically but not necessarily commercial or political) that employs euphemism, convenient vagueness, numbing repetition, and other such rhetorical subterfuges to create the impression that something has been said.")
Yea, I think it comes down to how different people weigh those qualities you lay out. I personally don't mind reading claudeish, but my tolerance varies on a few things.
Firstly, and maybe this is unfair, but if there isn't a disclosure about AI use, and it reads like an AI, my prior for "this information is going to be well-written and informative" drops substantially. AI admitted up front raises the odds that the author would also have read it and still published it (until I read enough things that violate this, too).
Second, I believe that I weigh the "well-written" part very highly, and have low tolerance for things that are vague/repetitive/incoherent metaphorically, and ones that feel just so *throat punchy*. Again, admittedly this is a vibe about the feel of the piece rather than the information.
I believe that sloppy writing leads to sloppy ideas, and in a book review I'm looking for an idea that is already fairly refined. I think someone else said it well: reading this feels like I have to stop and trace spaghetti back, and good "review" writing feels like the interweaving story lines flow sensibly.
Of course I also can overlook it if I'm familiar with the subject matter - easier to cut the spaghet.
For me it's a mix of i. and iii
The piece is decent, nothing I'd vote for but passable as a nominee and I finished reading it. It's also profoundly unfair to every human being (of which I am not one) who did painstaking work on their own entry, putting in far more time than an AI would because of the inherent limitations
And as a supporter of humanity in general, reading AI writing without knowing it genuinely feels like a violation. It's like I've been tricked into crossing a picket line.
Good analysis, thanks! I would agree with one of your points, that using AI in a contest that is specifically and exclusively for human writing is cheating (but of course using AI in a writing competition that permitted it would be perfectly acceptable!)
I couldn't agree with your second point - I think we need to judge things on their intrinsic merits and not on their provenance, and if the AI writing is so close to human that it's undetectable (or detectable only with special AI-powered tools) then I think the claim that it is inferior purely because it's AI is basically just chauvinism.
Of course this doesn't apply to this review (which apparently absolutely everybody except I could instantly tell was AI...!) it's more of a general philosophical opposition to ad hominem* arguments. I would defend human writers from a "your demographic is intrinsically not able to write about subject X therefore whatever you write is automatically inferior" charge using exactly the same philosophy.
(* I realise the irony of using the word "hominem" to include AI..!)
> 1. Is it AI-written, or AI-edited, or merely human written with AI advice? Is it even possible to tell? Does it matter which it is, or are they morally equivalent?
I do not think that it matters. I also still do not think it is a question of morality. Judged as an essay the review has a problem. That problem is very easy to pinpoint on "AI has written it" and very hard to pinpoint on specifics if we taboo that statement.
---
> 2. Are there any problems with the arguments/claims themselves? Are any of the excerpts hallucinated? Are any of the other claims about the book unsupported by the book itself?
I have not read the book but I am willing to charitably grant that the book is represented accurately and / or within the margin of interpretation we would allow a human reviewer. But what _is_ the thesis of the review so that we could diagnose problems with the thesis? I felt it was meandering without ever getting to any point. Saying nothing wrong by not saying much at all with many words.
---
> ii. Or is it because it's actually low-quality "slop" (as discussed above), and it's being slop is the real problem, not whether the slop was human- or AI-written?
Look, I know that this will sound esoteric and I would love to have a better model for my criticism but by first approximation I would say a human could not write this. As I have said elsewhere it is super easy to imitate free tier CharGPT but this is on a different level.
So I think "how would you judge this from a human" is a meaningless counterfactual. The text is inhuman in a somewhat subtle but, to many people, recognisable way, as proven by all the commenters recognising it.
---
> iii. Or is it purely because people hate everything that's AI written, regardless of its actual quality?
Hate is the wrong register. Everything is too absolute. When I use an AI, I obviously do not "hate what it writes", or I would not be using it.
I do not believe that pre Astra AI (I am hedging because I have not seen much Astra) can produce an interesting book review and I believer this largely because I think the review at hand is an earnest but failed attempt to do exactly that.
I predict that I will hate most book reviews written by pre Astra Gen AI and in a sense that is indeed because they are written by an AI but if next Gen were much better at writing like a human I would not reject that just because it is AI if that makes sense.
---
> 4. Do the opinions of A and B have merit? Are they coherent, consistent positions that argue a consistent case and present legitimate evidence for it? Or do they just give the illusion of doing this somehow (in a way that certainly fooled me!) when really the argument has no substance?
Again, what are the opinions they hold respectively? (Making this "adversarial" no less). What is the substance?
People have pointed out that B does not sound like an AI that much but B only seems to exist to push back on A without ever taking initiative.
I half expect the author of the review to come out an claim that they are B, A does not correspond to a human and all of this is an experiment in AI usage.
As mentioned elsewhere I am fundamentally uninterested in the "opinion" of a machine. The machine is useful to get a feeling for the average human opinion, some sort of common practice or lowest common denominator, you know what I mean. But that is the opposite of what a book review is supposed to provide.
My experience with AIs is that they are easy pushovers as soon as you voice strong opinions and why wouldn't they be? The AI has nothing it wants to convince you of. [I know this is controversial in this forum, oh well...]
Thanks for such a detailed reply!
> "what _is_ the thesis of the review so that we could diagnose problems with the thesis?"
I don't think it really has a single, central thesis... and for a book review that's perfectly fine? Sure, a *Scott Alexander* book review tends to have a sort-of bipartite structure, with the first half telling you what the book is like and the second half using the book to cleverly make some deep philosophical point that only makes sense because of the context established by the book - but I don't think this is de rigeur for book reviews in general?
So, instead of a single, central thesis, I think the book review is simply telling you what the book was like to read, from the perspective of one reader who liked it and one reader who didn't. Each reviewer makes some arguments for why they did/didn't like it, in support of their opinion of the author, etc., and the other reviewer sometimes responds to these arguments, but no firm conclusion is reached and both opinions are presented to the reader with roughly equal weight. I think I might feel short-changed if Scott were to write this sort of study - after all, Scott's conclusions and considered opinions and stuff are pretty much what I come here for! - but I don't think it is necessarily a bad thing for some other author(s) to write like this, if the subject of their study suits it.
> "a human could not write this"
Surely, for all the problems AI writing has, it acquired those problems by copying bad human writing? Or are you saying that AI, rather than merely copying humans, has invented a new unique writing style that humans couldn't copy? (I'm not trying to put words in your mouth here, so sorry if this is way off - I genuinely don't understand!)
Separately, in this section I wasn't asking whether the writing can be judged to be AI rather than human (I accept that it can, even if I apparently can't do it myself!) but whether, once we've established that some writing is "slop", it matters whether it's AI-produced slop or human-produced slop.
> "if next Gen were much better at writing like a human I would not reject that just because it is AI"
This is getting to the core of my question - thanks! I think I can break the question down further, though: do you need the AI to write like a human specifically, or just to write well in general regardless of whether the good writing is human-like? Or, do you consider that "writing well" and "writing like a human" are the same thing (perhaps because humans decide upon the aesthetic principles that count as "writing well")? Or, if the AI writes exactly like a human writer, only a bad one, would that still be interesting/worth reading for you?
> "what are the opinions they hold respectively?"
I shan't bore you by listing all the individual claims made, but here are a couple of random (heavily condensed) examples:
Reviewer A: "Kay's caustic tone is a result of burnout that's really common in the NHS, not his personal flaws". Evidence: "Registrars are fairly junior but have no top-cover"; "German/USA doctors have lots more support"; "specialists in other fields talk about patients the exact same way when they're under pressure and if anything Kay is on the gentler end"; etc.
Reviewer B: "Kay's burnout is caused by his own martyr complex and the NHS is merely capitalising on this opportunistically, not systematically overworking everybody all the time". Evidence: "Kay claims to be overworked but voluntarily spends hours with patients when he isn't asked to"; "Kay never asks colleagues for help"; "Key twiddled his thumbs for a week when expected patients didn't arrive"; "roster-makers really like to have martyrs available"; etc.
There are a whole bunch more opinions/claims like that, essentially!
> "As mentioned elsewhere I am fundamentally uninterested in the "opinion" of a machine"
I think this is the our main (perhaps our only!) disagreement! I would call this an ad-hominem, and say that to tell whether an opinion - or apparent opinion! - is good or bad we need to evaluate the arguments/evidence for and against it on their own merits, not just check who wrote it. But - this doesn't mean that I think AI text is necessarily worth reading! I think that most of it is terrible and if I notice that some piece is AI-written I often abandon it without really examining the opinion it is expressing. But this is just a heuristic based on experience, and when we want to know whether an opinion is good or bad rather than merely predict whether a piece is likely to be worth our time, we need to evaluate the actual evidence for that opinion.
> Surely, for all the problems AI writing has, it acquired those problems by copying bad human writing? Or are you saying that AI, rather than merely copying humans, has invented a new unique writing style that humans couldn't copy? (I'm not trying to put words in your mouth here, so sorry if this is way off - I genuinely don't understand!)
> Separately, in this section I wasn't asking whether the writing can be judged to be AI rather than human (I accept that it can, even if I apparently can't do it myself!) but whether, once we've established that some writing is "slop", it matters whether it's AI-produced slop or human-produced slop.
I believe that the AI that wrote this review is unable to asssign meaning to the words it outputs and that this, more than any stylistic "tell" unmasks it as an AI. [Again, I know this is controversial in this "AGI is near" region of the internet]. I would not know how to emulate this review - that is how to provoke the effect that this review apparently had on many readers.
This, as mentioned several times, explicitly in contrast to the stylistic shortcomings of other models. ChatGPT I find very easy to emulate. Maybe I am just unfamiliar with the reviewing AI's house style.
So what I am saying is the lack of meaning makes this "slop" and definitionally humans cannot produce text devoid of all human meaning. They can produce kitch or camp or a simple cash-grab, or subpar prose but not, by this definition, "slop". So the question is unanswerable. [Note however that definitions are arbitrary so this is not a claim of a deeper truth].
> Or, do you consider that "writing well" and "writing like a human" are the same thing (perhaps because humans decide upon the aesthetic principles that count as "writing well")?
Gun to my head, yes.
> I would call this an ad-hominem, and say that to tell whether an opinion - or apparent opinion! - is good or bad we need to evaluate the arguments/evidence for and against it on their own merits, not just check who wrote it.
Ad-hominem, I think, means directed at the person, and, is my point, there is no person here to dirrect criticism at or engage with critically. This is dangerously close to word-play but how do you ask an AI for a subjective opinion on a book if there is no subject to have an opinion.
I have a suspicion that we use "opinion" differently. If something can be backed up by argument / evidence, as you suggest, would that not make it a factual statement as opposed to an opinion? I mentioned this above but to repeat: I use AIs and I would not be doing it if I were uninterested in any AI output whatsoever. I am specifiacally not interested in _opinion_ from an AI.
Maybe not coincidentally, the opinions you chose as examples concern the relationship of the author to the system that "produced" him. This reading takes for granted that we are indeed reading a report of factual events.
But that implies a neutrality that is oppositional to any review of the book. What is there to interrogate in a neutral text? Even in non-fiction [and the book, without having read it, firmly seems to be fiction] there is a non-neutral choice in what information to present and how.
B (the maybe human) lampshades this: "Kay’s conceit is that these entries were written contemporaneously as self-reflection exercises and simply published later, but that seems unlikely to me." But then the review gets nowhere with that point. We can call that a weakness, AI or not. As an AI would put it: The fluency of the prose writes a check that the the flow of the argument does not cash in.
Long story short, I think the genre of a book review (as opposed to, say, technical dcumentation) is spectacularly ill-suited to be approached with present day AI.
EDIT:
Forgot to point out the sweet sweet irony that some people are defending the review with "death of the author" (as in authorship does not matter) whereas the review it self is very very interested in constructing a pyschogram of the author of the book under review.
Again, thanks for the detailed reply.
I think I get your point, now - rather than a specific writing style, made up of identifiable elements such as tone of voice, vocabulary, the expressions employed, etc., there's a certain undefinable je ne sais quoi to AI writing caused by the AI's having no internal experience, and this A) makes its opinions essentially illusory because you can't have opinions without internal experience, and B) can't be imitated by humans who have internal experience and can't simply imitate the je ne sais quoi because it's not comprised of identifiable elements? If so, I'm sorry that you had to labour the point so much to get me there!
On the "humans can't imitate the no-internal-experience style" idea, I don't have anything to say that's cleverer or more useful than all the philosophy written about whether one could tell a human from a p-zombie or not. Personally I think that a human could imitate a p-zombie but not the other way around, since the human has all the faculties the p-zombie has plus a few extra ones it doesn't have (eg. self-awareness, introspection, etc.) - but I admit that philosophically I'm a bit out of my depth with this sort of thing.
> "Ad-hominem, I think, means directed at the person, and, is my point, there is no person here to dirrect criticism at or engage with critically"
Yes, I understand that "ad hominem" means "against the human", but I deduce from it's being in Latin that it must pre-date LLMs and the expression-coiners probably didn't factor-in entities other than humans who might state arguments. Going with the sense of the ad hominem idea rather than taking the Latin name literally, an ad hominem is an argument where you attack (or praise!) the entity making the argument rather than the logic or evidence for/against the argument itself. So saying "This opinion is fake/meaningless because it is based on purported facts that are untrue" would not be an ad hominem, but "This opinion is fake/meaningless because it was written by an AI and AI can't have opinions" would be an ad hominem.
Note that Graham's Hierarchy of Disagreement, 2008, specifically defines an ad hominem this way - "attacks the characteristics of the writer without addressing the substance of the argument" - without requiring that the writer be a human: https://en.wikipedia.org/wiki/Paul_Graham_(programmer)#/media/File:Graham's_Hierarchy_of_Disagreement-en.svg
> "how do you ask an AI for a subjective opinion on a book if there is no subject to have an opinion"
The same way I would ask for AI for an answer to anything else! If the AI can use its encoded rules for how maths works* to answer novel maths questions I don't see why it couldn't use its encoded rules about how medicine works to answer novel questions about medicine (or any topic, including literary criticism, or the nature of the human experience). In every case, for the purposes of judging the quality of the answer it doesn't matter to me whether the answer given "feels like anything" from the inside, whether the AI truly believes it, or whatever - what matters is, does the answer constitute some true fact about the world that we didn't know before.
(* Scrupulously avoiding saying "knowledge", "understanding", etc. here! cf. https://slatestarcodex.com/2019/02/28/meaningful )
> "I have a suspicion that we use "opinion" differently. If something can be backed up by argument / evidence, as you suggest, would that not make it a factual statement as opposed to an opinion?"
Both! It would be an opinion /about/ objective facts. Some opinions do tend to be about objective facts - medical opinions being a great example! Where a medical opinion is good it accurately captures/represents the underlying facts and the opinion thus matches reality, and where the medical opinion is bad it mistakes or misrepresents the underlying facts and thus doesn't match reality.
I think the question of whether *all* opinions are about underlying objective facts is open - and is perhaps even more complicated and hotly-contested in philosophy than the p-zombies I mentioned above! Personally I would say that even if I had an an opinion like "faeries are cuter than unicorns" it would still be underpinned by objective facts about the world, just that those facts would happen to be about the makeup of my brain and not about frolicking magical creatures - but I am sure a philosophical dualist or whatever would disagree, and once again I'm probably out of my depth here.
> "What is there to interrogate in a neutral text?"
I think that most of the claims made in the review (of which I listed two, but of course there were dozens of such) can be examined and in principle tested against reality, and that for this testing it doesn't matter whether the claims were made by a human or an AI - we test them by looking at reality just the same. If the claim were etched onto a rock I would still want to judge it based on the evidence and arguments for and against it and I wouldn't consider "rocks can't have opinions" as sufficient grounds to dismiss it.
> "sweet sweet irony that some people are defending the review with "death of the author" (as in authorship does not matter) whereas the review it self is very very interested in constructing a pyschogram of the author"
Haha wow, this is interesting! I never noticed that! (Just to be clear, I don't think it's quite true: I don't think "we must evaluate arguments based on their substance rather than their authorship" reduces to "authorship doesn't matter" - but it is very amusing to notice nevertheless!)
I think it makes for a useful example, too: where Simon Kay makes claims about the NHS (eg. "Registrars have no top-cover"), the reviewers assess these claims with regard to their own observations about the world ("At 0300 the consultant is at home") and with regard to Kay's character/personality ("Kay never asks anybody for help anyway").
I believe that "Kay never asks anybody for help anyway" is an ad hominem and, though it might constitute extremely weak evidence for the proposition "Registrars have no top-cover" in a Bayesian sense*, it is essentially useless for evaluating that proposition whereas the "At 0300 the consultant is at home" observation is pretty useful for evaluating it.
(* /Yet again/, though, I'm probably out of my depth here...)
I don't have much more to add but I want to clarify that I am _not_ going into this with the axiom that LLMs never could develop meaningful understanding because they lack a soul or something like that.
The other way round: lack of meaningful understanding is my best guess as for why this particular LLM's writing feels so alien to many people including me.
Others in the comments have been pointing out that they can see direct tells that I seem to have missed (as opposed to a vague feeling of alienation). But, to loop back to your original question, something about the review signals "AI writing" as opposed to "bad writing but human". What is it?
On a more practical level, if the big labs could "beat" Pangram, would they not be doing it?
> 3. Why does everybody hate it?
A little bit of both.
Imagine if someone hired a human ghostwriter to write their essay. You start reading it and immediately think, "wait a minute, was this written by Douglas Coupland?"
Does this warrant a disqualification? Maybe. It depends on the rules of the competition.
Does this mean the essay itself is bad? Again, it depends. Maybe you love Coupland's style, maybe you hate it. Maybe you're fine with it in moderation, but not when every other piece of copy sounds like it came from under Coupland's pen.
LLM-written prose is very distinctive and uses the same rhetorical tricks over and over again. You buy a CD by a new band, and it sounds like Nickelback.
I for one am happy to call out writers using AI when they should not, and in that sense I suppose you can say I am witch-hunting. It's because I think it is bad for humanity for there to be a norm that it is OK to use AI to write for you and pass it off as your own, and this is the only moment at which there might be some chance to establish that (I believe important) norm.
I wouldn't say that was witch-hunting! I admit I don't really know what "call out" means (surely you cannot literally mean that you demand the AI-user steps outside their building so as to duel with you to the death...) but if it just means some reasonable response like "vote against" or whatever - that seems to be the system working precisely as intended, rather than a witch-hunt? You're allowed to vote against reviews for whatever reason you like!
Haha, no, I don't challenge anyone to a duel to the death. But I might publicly say, "Why did you write this with AI?" even if it causes social discomfort.
I try not to let contemptuous, morally injured individuals anywhere near my vagina. Especially during the most mentally, emotionally, and physically intense and vulnerable activity of my life - childbirth.
Kay might not be specifically misogynistic so much as generally prickly. But even so, birth is an activity whose success and safety depends in part on keeping my oxytocin high. Oxytocin is the warm fuzzy hormone people feel after exchanging a safe hug, and it is also the hormone that powers contractions and prevents uterine hemorrhage. So letting hostile, oxytocin-undermining people into my space would negatively impact actual clinical outcomes.
I wonder if midwives have a better experience in the NHS than OBs do. Certainly the statistics in the UK, as well as the NL and US, show that for low-risk mothers and babies, home birth with a midwife is far safer than hospital birth with a midwife, which is in turn safer than hospital birth with an OB.
How do you square this reading with the "She and her husband..." excerpt? Do you see Kay as an unpredictable, Jekyll and Hyde type figure; caring and supportive at some times and contemptuous at other times? Or do you see him as contemptuous throughout and consider this lone excerpt simply not representative of his manner? (If so, why the anomaly? Is he simply trying to make himself look good in this one passage?) Or are you less certain about his bedside manner than you appear - but because it's literally a matter of life and death for you and your baby you could never take a chance and must assume the worst? Or something else entirely?
Having not read the book myself, I can't say which manner is more typical of him throughout the book, much less his actual practice. But yes, based on this book review, he does seem to be rather Jekyll and Hyde. That's not a risk I prefer to take if there are more reliably humane attendants available for the most important day of my and my baby's life.
Kay did well by the parents of the stillborn baby. His bedside manner was just what they needed, apparently: "Stick to what you know. I just talk practically about what will happen over the next few hours. They have a thousand questions, which I answer as best I can. This is clearly their way of coping for now, medicalizing it."
That's similar to what I might want in the weeks leading up to a surgery or other medical procedure. I did have a minor surgery once, and my OBGYN was helpfully informative. She was also kind, which I liked, but I wouldn't have been traumatized if she had just been informative and not kind. I had questions, I got answers, I made decisions. Then my job was done, I went unconscious, and she did her job. That's surgery. I'm glad OBGYNs exist, and I have no reason to doubt Kay was a competent surgeon.
OBGYN is a tricky (and for some, excitingly well-rounded) field, because it is taught within the surgical tradition, where, famously, doctors prefer their patients unconscious; but via regulatory capture it also now engulfs aspects of midwifery. And those are two very different bedside manner skill sets.
In the passage about the stillbirth, I see Kay being a kind and informative surgeon. But what he offered them - having a lengthy conversation that medicalizes the experience - still isn't typically the needed bedside manner during normal birth.
Yet even a lengthy Q&A with any birthing mother who was able to hold a conversation would probably have reduced the unnecessary C-sections he admits to ordering when he was suffering from PTSD.
I'm not trying to argue that Kay was worse than the average NHS OB. I don't have data for that. From the review, it sounds like the NHS makes many OBs worse than they have to be, which is the central travesty here.
I do know that if I were Kay's patient and were planning a normal birth with him, and I got a whiff of any red flags of contempt, moral injury, hostility, or even prickliness that I wouldn't mind in other contexts, I would switch providers quickly. I would switch birth providers much more quickly than I would switch other medical providers for something as personal as "bad vibes." If they kneecap my confidence in their office, they could do the same in my birth space. And since birth is my job and evolved to partially run on vibes (ok, interpersonally influenced hormones), that actually matters.
Thorough and well-argued. Thanks!
The unnecessary C-sections *are* the concerning part, and I do have to wonder how many patients (and nurses, etc.) had their objections ridden rough-shod over, with him invoking the usual "I am The Doctor here, that makes me God" authority to get his way.
That is him being contemptuous of his patients' fears and wishes, in order to make himself feel better. In a very understandable way! We get why he was so paranoid and controlling! But it doesn't make the allegations of misogyny go away, either.
Oh dear, I think that might be the most uncharitable interpretation possible!
If we take the text at face value (which I grant you we can't always do, necessarily!) I think it's vastly more likely that he was ordering unnecessary C-sections because he honestly believed that they would be (marginally) safer for the babies and he absolutely did not want to expose any more babies to even the slightest risk, not because he was being contemptuous of parents' wishes or getting a kick out of exercising his authority. I think it's vastly more likely that he was motivated primarily by fear for the babies' safety than by a desire to merely make himself feel better. And I don't see any reason to think that any of this was misogynist or gender-related at all!
As for the "controlling", perhaps I don't understand well enough how such decisions are made, but my impression is that a doctor simply reviews each case, decides whether a C-section is necessary, and just tells everybody that in their medical opinion this is the safest course, not that the doctor has to argue vehemently for their position and harangue skeptical parents, nurses, et. al. into acceding, Dr-House-style? If so, I don't think it's controlling for the doctor to say "In my medical opinion we need to do a C-section" instead of "In my medical opinion we should not do a C-section". (If however it does actually go the Dr House way then of course I would agree that it would be controlling for Kay to insist upon the C-sections over everybody else's objections!) Either way, the obvious mitigation for the problem of doctors prescribing the wrong interventions would seem to be a medical second opinion, but I don't think the NHS has the resources to provide that for every single case - and I think this is one of the problems Kay highlights throughout the book.
(nb. needless-to-say I fully agree he was wrong to order the C-sections, and that people with PTSD aren't qualified to assess the risks properly and should not be placed in a position where they're obliged to make such decisions. I just see Kay's being in that position and consequently making terrible decisions as a (major) institutional failure, not some character flaw of his)
The point is, if the operations are unnecessary, then he has to convince the women to have them. He has to convince the other hospital staff that they are needed. And by the sound of it, he was so paranoid that he railroaded these through. I'm not confident that if a pregnant woman said "No, I don't want one" that he would take no for an answer, given the general tenor of the quotes.
I think the core of my disagreement here might be your "has to convince". What does that entail? As I asked before, is it more like the doctor mostly saying "In my medical opinion you need to do procedure X" and the patients/staff/whomever mostly saying "Okay doc"? Or is it more like an episode of Dr House with much haranguing and bitter arguments all around before the doctor's opinion, correct or incorrect, is eventually accepted with great reluctance? In my not-inconsiderable experience of the NHS it's much more typically the former - which I agree is a dangerously flawed system but not that it's "controlling" or "contemptuous" - although of course I have to admit that my experience doesn't include obstetrics specifically.
> I'm not confident that if a pregnant woman said "No, I don't want one" that he would take no for an answer, given the general tenor of the quotes
I'm struggling to see what quotes you mean, here!
Kay's completely disregarding his consultant's instructions to deliver the baby he was ordered not to deliver? That seems like an example of good character and patient-focused care, to me, not a God-complex!
The caustic tone with which Kay describes some of his more socially embarrassing patients and their problems (which reviewer A described as something like "on the gentler end of the spectrum for the medical profession generally")?
The episode where he heavily medicalised all the patient's problems rather then expressing sympathy? I agree that may not have been what that patient needed (it's exactly what I want from my physicians but I admit it's probably not what most people want!) but I think the fact that he stayed until beyond midnight (after a supposed 18000 finish) in order to support this patient indicates that his heart was in the right place and he was doing his best for them, here, not trying to deliberately give them what he wanted rather than what they need.
Finally - even if you're completely correct that he wouldn't take no for an answer, I think it does matter *why*: if he is genuinely doing his best to advocate for the baby's safety and trying to protect it to the best of his ability I think that suggests a very different personality to if he is contemptuous of the patient's wishes and mostly just cares about "making himself feel better"!
Having read the book, I think Kay’s general attitude was paternalistic even before the disaster that made him super-cautious - he preferred to make the decisions rather than presenting patients with the options and letting them decide.
Can parents-to-be in the UK "switch providers" the way you suggest?
I've never done it myself but I'm pretty sure you can! The UK has a two-tier system, with both nationalised and private healthcare. If you use the nationalised healthcare I imagine it's not very easy to change hospital or whatever - but if you use private healthcare you can choose whatever doctor and hospital you like!
(Obviously private healthcare is shockingly expensive - but I think it is still cheaper than USA healthcare for many treatments, thanks to the need to compete with the NHS...)
I thought private medical options in the UK were easy for cosmetic surgery and other electives, but quite difficult and sometimes completely unavailable for things like emergency medicine or obgyn? Maybe I have been reading the wrong sources. Traditional media reporting has many stories about UK families flying to other countries to access treatment options they couldn't otherwise access at all.
Again I don't know from personal experience, but I think you can have most procedures done privately? For example, here's UCL's (a prestigious London teaching hospital's) price list for obstetrics. A C-section seems to cost about £8500: https://www.uclhprivatehealthcare.co.uk/services/maternity/private-maternity-prices
And here's a London-based provider (admittedly one that I've never heard of before; they were just the top DDG search result..) offering private Urgent Care: https://www.hcahealthcare.co.uk/locations/urgent-care-centres
Obviously if you get run over by a lorry whilst crossing the street you'll automatically get taken to the nearest (NHS) A&E department, but if you're in a well enough condition to choose for yourself where you go (and you're willing to pay..) I'm pretty sure that you can go wherever you like, and that there are private providers for most procedures.
The NHS covers home birth midwifery care, hospital-associated midwife-led birth centres, and obstetric units.
You can switch NHS providers and/or birthing locations at any point during pregnancy or labor, subject to availability.
Interesting - the stats in seeing (for the US at least) show a significant increase in risk for the child during home births with a midwife. Rough stats seen to show an approximately 3-4x increase above the admittedly base rate set in hospitals.
I'm also concerned there could be a sampling bias here - if you know you're going to have, or even suspect you might have, a high risk pregnancy there's almost no chance you'd or for a home pregnancy. Any comparison of right is necessarily going to run into the fact that I've pool is a general sample and one has been selected to be low risk.
May I ask which source you're looking at?
My recollection, though it's been years since I looked at the actual studies, is that although babies have somewhat poorer outcomes at home for first births, they have slightly better outcomes at home for later births, compared to hospitals, which averages out across both cohorts to about the same. And mothers overall have far better outcomes at home than in hospitals.
What I've seen has all been comparing low-risk births only, so there's no sampling bias from that. Hospitals are uncontroversially the safer place for high-risk births, such as twins or, as in the example in the book, placenta praevia.
This is also only including home births that are attended by a qualified midwife (not unattended "freebirth").
And it compares planned home birth (including transfers to hospital if something goes wrong) to planned hospital birth (including accidental home birth if baby comes quickly).
>Certainly the statistics in the UK, as well as the NL and US, show that for low-risk mothers and babies, home birth with a midwife is far safer than hospital birth with a midwife, which is in turn safer than hospital birth with an OB.
I'm afraid I think this is dangerous misinformation - home births are significantly more dangerous than hospital births.
It may be that you're being mislead by selection bias - at-risk mothers are disproportionately likely to opt for hospital births and have doctors present.
What I'm referring to only compares low-risk births in hospital vs. at home, so there is no selection bias on that factor.
To be sure, high-risk births are far safer to approach in hospital.
That's not how selection bias works. At most, you could say you attempted to control for the bias.
I'm not sure what you mean. Let's say I run a study comparing patients with a simple fracture in the ulna or radius (the forearm bones), to see whether they have better outcomes if they go to a green-painted or blue-painted emergency room. The patients choose where to go, but their fractures are equivalent. I exclude patients with compound fractures, which everyone thinks are treated better in blue-painted emergency rooms. Have I controlled for selection bias? Would sampling bias be a more precise term for what's actually being controlled for?
The key issue here is your assumption, "their fractures are equivalent." A "low-risk" or "high-risk" birth is a crude binary classification that cannot hope to account for all preexisting medical differences. Restricting your analysis to "low-risk" doesn't fully answer the question of, "Are mothers with higher-risk pregnancies self-selecting to hospital deliveries?"
I don't want to be rude here, but this is extremely obvious and I am confused by why this is not extremely obvious to you.
Ok, now I see what you mean. Yes, it's possible that could be happening.
I do spend a lot of time with mothers who choose each location for their birth. In my experience, people who prefer home birth have a strong enough preference that they do not tend to opt for the hospital instead unless they truly "risk out" of home birth care such that the midwife is not willing or legally allowed to take them on as a client. But I don't know everyone.
Regarding the future of the NHS mentioned at the end of the review, I think there is actually a good case for not worrying at all. Either the singularity happens, and we have never-tiring never- rude always-sympathetic all-knowing all-seeing super robot doctors, and therefore we don't have to worry about the NHS; or the singularity happens, AI decides to kill us all, and therefore we also would not have to worry about the NHS.
Would people be less upset about artificial doctors and surgeons than they seem to be about artificial authors?
Why would they be allowed to be upset? That would cause unnecessary suffering.
Or the singularity doesn't happen, and we stagger around in a haze of confusion as always.
This was way better before I spend two months reading Opus 5 outputs at work huh.
What terrible job demands that?
Software engineer.
When A defended Kay from the charge of misogyny I thought, 'They must not be familiar with his work with Amateur Transplants.' Then B mentioned Amateur Transplants!
To be clear, I like Amateur Transplants. But even if you'll give a gynaecological pass to "The Menstrual Rag", I don't think you can listen to "Nothing At All", "Northern Birds", "Sheila's Wheels", "New Man Song", and "Couples Counselling B" and be like 'Well I'm sure no one involved here was a misogynist.'
You asses spoiled it god damn it why did I go into the comments before finishing god damn it.
The AI bothered me, especially in some paragraphs, but overall and seemingly against popular opinion I did not dislike the review, also because the book seems very interesting. I like the two-author dialogue style, personally.
S ojm
Mo lnljl.$. So v vs Ben vs
Bnspziz Jinnah is as
Just to save the next person the trouble, I can confirm that this is not rot13.
I cannot agree more! Give Jinnah a ham sandwich and a shot of the finest gin!
Very strange to get to the end of the piece and find this conclusion:
"The honest answer about the NHS is the one nobody campaigning for it wants to give. The system as currently configured is not sustainable. The routes back to sustainability all involve pain. Either austerity acute enough to cause real ischaemia, then amputation of what falls off, or eventual collapse, which will hurt more, and at a time of nobody’s choosing."
Isn't the whole piece about the disastrous effects austerity has already had on the NHS? The obvious solution this points to is reversing that austerity - paying doctors better than they were paid in 2008, spreading out the work among more people, investing in all the systems that are deteriorating, etc.
where is the money coming from?
A modest suggestion is clearly forming in my head, reading studies like these: https://acss.org.uk/wp-content/uploads//Wealth-inequality-and-growth-in-the-UK-policy-briefing-Nov-2025.pdf
It's same the thing happening all over western europe.
See the previous paragraphs about the British populace banging pots for healthcare workers and then rejecting plans for pay increases 8 months later.
I think the writers take it for granted that actually finding nhs isn't going to happen, and that austerity will be the result.
The article doesn't actually establish that "the British populace" rejected pay increases, it just mentions that certain newspapers took a harder editorial line. Apparently a poll this month said "78% of UK voters support increasing overall NHS funding, even if this requires tax rises."
I find this quote to be a pretty good summation of his message on this point:
"The public’s affection for the NHS, the abstraction, turned out to be largely separable from any practical commitment to the people who staff it. It is possible to clap for someone in March and curse them in October without acknowledging any contradiction, and this capacity has been demonstrated at scale."
Yes, but he doesn't seem to actually mention the *public* cursing junior doctors for wanting better pay/conditions. He says much of the *media* did this.
I wasn't following this too closely at the time, but it would be EXTEMELY on-brand for the UK tabloid media to make a big deal about clapping for medical staff during lockdown ("Everyday heroes! We're all in this together! Dunkirk spirit!") and then demonize them a few years later when they ask for things that would cost taxpayers money, especially when the term "junior doctors" makes it easy to spin a narrative of lazy, entitled young people.
the public were not asked, that decision was made by representatives.
Austerity could mean higher taxes, or cutting spending other than on the NHS.
Austerity means lower taxes, or at least tries to imply it. From the above context, I assume it here refers to cutting some services provided by the NHS to improve others.
Austerity just means lower spending.
Yeah, this is what I'm saying.
It’s not what you are saying. You said lower taxes. Austerity almost always is both higher taxes and lower spending.
I said "lower taxes or tries to imply it". Whatever the government ends up doing, it tends to get sold to the country as "we're saving you money".
(In some cases, it's sold as "we're spending it on more important things instead", but you don't do that with the NHS. Telling Brits the NHS isn't important doesn't go down well.)
No, austerity does NOT mean "lower taxes". When the IMF makes countries adopt austerity policies, it typically involves things like hikes to VATs that result in taxes as a higher percent of GDP.
Let me rephrase: in the UK, when the term "austerity" is applied to the NHS or any other public service, it means cutting state funding of that thing. It might be used to mean something different in the USA or wherever you are?
The USA doesn't need loans from the IMF. But I can see how the term "austerity" applied to a specific department, rather than the government as a whole, could mean that because the NHS can't levy taxes by itself.
I think the term "austerity" might have a technical meaning among economists that I don't know about? In UK media, it just means cutting funding from a thing. "NHS Austerity" means less tax spending on the NHS, "Welfare Austerity" means cutting benefit taxes. And so on. It doesn't describe your current relationship with the IMF.
It's AI prose, don't be surprised when the colourful language fails to cohere into meaning.
I enjoyed reading the review, and did not notice any AI issues until reading the comments.
One thing I did find a bit jarring was the quote "But you said there were three reasons the book was good, and I only count two" which seemed unlikely to happen in an actual adversarial collaboration.
Unfortunately this is so obviously written by an LLM that I couldn't finish it
Depressing to see Claudeslop in the book review finals. Once you know it, it’s extremely grating, and the fact that AI writing made it to the finals makes me very sad.
This is the first Book Review that I can't make it more than a few paragraphs into. A rare miss in the contest series. Hopefully next week's is back to the usual high bar.
I really enjoyed this review, and didn't notice any particular AI abominations in the writing, but I'm not particularly sensitive to that kind of thing.
I enjoyed A's commentary and appreciated him clearly using this text to express an issue he was working through himself. Kind of would have appreciated a bit more detail about what caused him to quit, but I get that's somewhat veering outside the confines of the book review.
I'm honestly confused what B thought he was adding to this review though. B's argument seemed to be that Kay was never for to be a doctor and that his eventual collapse was always going to happen, but he never quite explained why he felt this. I'm struggling to explain this, but his illumination of Kay's flaws (martyr complex, cynicism) were never substantive enough to bring me over to a point where I agreed Kay shouldn't have been a doctor.
B also kind of got bowled over throughout this, which read weird in what was clearly an edited piece. You'd have thunk he'd realize his arguments were under baked and would have added some more.
Oh man. I really liked this review and didn't expect that it was AI. Maybe this makes me a bad person, but as long as the thoughts were actually from A and were just "polished" by AI, and not made by AI whole cloth, I'm pretty ok with it. I guess the issue is knowing which case I'm dealing with.
Same. If there actually is an NHS psychiatry resident with the ancestry and life experience described who prompted AI to write this review and signed off on the result as faithfully representing their own perspective, I'm fine with that. It's glib in places and isn't going to win any awards for concision, but it's gripping and interesting nonetheless.
But if the whole persona of A is just made up, if the review leads us to believe that a particular insider to the system has this perspective on the book when in fact there is no such person, then *that* fraud, not the use of AI as a writing tool, should be disqualifying.
I think my bar is a bit higher - I'd want them to put the ideas down and use Ai like an editor. But yeah it's a spectrum and definitely hard to know how it was done...
I doubt A is entirely made up, because there are some lines attributed to them that Claude would never have written. For example,
"Bevan, who had a sharper sense of professional psychology than the BMA gave him credit for, executed one of the great political manoeuvres of the twentieth century. In his own words, he 'stuffed their mouths with gold.'"
Claude typically doesn't pass moral judgment on a person or repeat their quotes verbatim unless explicitly instructed.
I reacted with "this is AI" from the very first sentence (not counting the quoted material).
I think Scott is including this as a test to see whether AI can write book reviews his audience will like.
Since liking comments is disabled, I will simply add another comment saying I dropped out of this essay a few paragraphs in due to obvious AI writing.
I rather like the ACX Tweaks extension. Since only a minority use it, there are a lot fewer comment likes than there used to be, but I think it is worth installing.
Enjoyed the review, scrolled down hoping for some interesting discussion, but just found people talking about whether it was AI-generated. Sad! (At the risk of throwing more fuel on the fire, the AI thing never occurred to me while I was reading it.)
Many people are saying, indeed. I keep seeing these types of statements around Substack that the NHS is failing due to governmental austerity, then mention it to irl friends (some of whom are actual Britons) and they're like no actually it's Evil Corpos trying to Privatise All The Things via Regulatory Capture. Which I guess is not strictly incompatible as a root cause? But suggests a very different diagnosis than just throwing more money at the programme. Would have been great to see some SME discussion along those lines. Heck, you're Australian as I recall, I'd be interested to hear if "NHS Expats Defect To Down Under Docs" is an accurate summation! But instead it's relitigating Claude v. Readership all the way down.
Which proves the point, I think: imagine the fruitful discussion that could have been had, if this review hadn't been (probably) AI-written! In some sense it won't matter how much AI actually speeds up and improves the process of writing, if people Semantic Stopsign the resulting output anyway. And there's some collatoral grammatical damage being done as well, where a whole swathe of perfectly good verbal tics and rhetorical tricks fall into disfavour just because they're associated with LLM writing (of the time). Pour one out for the disgraced emdash, and all that.
The NHS, as with English schools, is suffering in large part because everything else had been cut, and so they are the only things left to deal with every societal ill. That’s where austerity really comes in, although the pay freezes weren’t great.
The NHS has the added problem of an aging population and a lack of social care, meaning that many hospital beds are occupied with elderly people who aren’t _refusing_ to leave as the review weirdly suggests (this must surely be an AI contribution), but cannot leave as there is nowhere safe for them to be discharged to. They don’t need hospital care, but they are not firm and healthy enough to go home as is. This is why people keep trying to bring up social care, and how to square spending on that really is the eternal struggle that is likely to destroy the country.
On top of that, there have been various bizarre decisions made about NHS hiring practices and training pipelines which have had inevitable long term consequences, some of which end up in hot topics and so nobody ever discusses them properly. The review might reasonably have been about that, but mostly was not.
Huh, I remember some time ago hearing of a similar phenomenon in Japan...except there it was seniors "malingering" in the justice system rather than the healthcare system. Nowhere safe to go home to, at least in jail you've got a bed and meals and someone to check up on you. In both cases this seems like a deeply distorted and unsustainable state of affairs. Makes a lot more sense than "bed blockers" too - people everywhere complain about getting discharged before they feel perfectly healthy, but it'd be really strange if they could insist on staying longer and regularly get their way, not unless it was in private care or otherwise paying out the nose.
The Black Wednesday thing in particular sounds like an insane training practice. I know of a retail company that likewise has a habit of transferring managers on a dime, with a large geographic net...but there it's not even coordinated on one specific date, it's just arbitrary and sometimes there's only like a week's notice given. Predictably this results in a lot less interest in getting promoted, and selects for a particular kind of tumbleweed. It's hard to just uproot your life like that!
> Heck, you're Australian as I recall, I'd be interested to hear if "NHS Expats Defect To Down Under Docs" is an accurate summation!
There's definitely a lot of British, and also Irish and New Zealandish, doctors and nurses in the Australian system. I spent a few days in hospital somewhat recently and actually chatted with a few nurses from these countries who talked about how broken things were at home and how happy they were to be here now. (I think the money was a big factor too, and it's also worth noting that there's almost no Americans, who also like to complain about their broken health care system but get paid a lot more at home.)
I don't know whether the Australian healthcare system is genuinely well designed, or whether we're just a decade or so behind in the collapse. The same structural problems of an aging population needing ever-increasing amounts of health care, and the constant development of brand new expensive treatments for previously-fatal conditions, are the same everywhere.
Wonder how much is just the weather. Plenty of different barometers for different folks in the USA, not so much in the microclimes of an island nation. Less of a language and culture lift too. And, yeah, one mostly hears of medical migration into America, not out of it.
Demographics did end up being destiny, just not in the way that phrase was originally used. Happy to have reached mid-30s with several friends and relatives my age ending up having kids. One very tiny contribution towards righting the pyramid. (But of course at the other end, I've got multiple relatives in the 80s-90s, and at least one who's made it past 100. The sums tossed around when discussing their care...it's a dificult problem.)
On the review- I guess it's reassuring to know that despite all the progress, there's still a text-based skill that LLMs are clearly inferior to humans at performing. Almost unreadable by the end.
On the book itself- as a former NHS junior doctor (left before we became residents!), I always felt like This Is Going To Hurt was less of a literal memoir/diary and more like a stand-up routine, where Adam Kay took a collection of things that had happened to him, had happened to other people, almost happened, or would have been amusing if they did happen, and then turned them into a coherent narrative without ruining the fun by disclosing that. I'd be very surprised if every thing that was contained in the book genuinely happened to him personally, as described.
House of God is a much, much better book. For a less laugh-out-loud but probably more accurate representation of junior doctoring quite a few years before TIGTH, Max Pemberton's "Trust Me I'm A (Junior) Doctor" is worth a read. Otherwise, the book is great if you read it for what it is, which is very funny, rather than expecting anyone's day job to actually be that funny/interesting/dramatic.
I read the whole thing. I was also very struck by how much it sounded like Claude (which was bad). I thought it more likely than not that it was by Claude, with the other possibility (as Thomas Johnson raised) being that people are starting to write like Claude.
Glad it was the former.
I didn't enjoy the review very much. Parts of it were interesting. Parts of it were pretty incoherent.
For example:
"The dominant reading frames Kay’s exit as a sane person escaping a mad system. I endorse something closer to: a man with a treatable post-traumatic syndrome whose only available coping strategy was career exit, because the system that broke him had no apparatus for putting him back together."
He then goes on to talk at length about how much more damning the second option would be for the system. Really? A broken system not having a way to fix somebody it damaged is worse than a system being "mad"?
We have "Psych textbooks love to blabber on about various kinds of defense mechanisms, and humour ranks high on the list." and then later, "Any psych textbook will yap on about mature defense mechanisms, of which humor ranks highly." This point is so amazing that it gets a third mention as well!
Then there's the "range of registers" point. We're told halfway through that Kay can "be funny in at least four distinct modes." Three of them are ennumerated, but the fourth, we're told, "And there is the tender register, which is rarer and harder to spot but is what elevates the book above a string of anecdotes. I’ll save the strongest example of this for later in the review, where it makes for the best argument."
Later, when he brings out the "saved up" passage, we find it's a harrowing story of an infant dying and Kay crying over the corpse. Certainly affecting, and powerful writing (by Kay), but let's recall these were supposed to be distinct modes of *being funny.* Ummmm?
These are the kind of larger-scale missteps that caused me not to track with A's case. Of course, even enumerating them feels a little like arguing in public with Claude, but I figured it might be useful. The more obvious annoying parts were the AI writing at the "local" level of paragraphs.
I will say that I enjoyed B's voice and writing and largely found his points persuasive. I suspect he could have written a good review of the book.
Finally, I'm very sorry to see that A is struggling so badly with depression that he "almost succumbed." I assume this is true and not AI generated. Writer A: I'm very glad you did not succumb. Please hang in there. DMs are always open. Probably your bad experiences will make you a better psychiatrist in the long term. But if that sounds awful, or if doing it there in the UK sounds awful, there are other options that are also good. Don't let the broken system become your whole world.
(Because several people said they didn't notice any very AI-ish writing, I'll share a few of the many that jumped out at me, although even less egregious ones also felt suspect.)
"a self-indulgent memoir whose author offloads his exhaustion onto the bodies of female patients, dressed up in jokes that are funnier to him than to the women they reference"
"The book is good for at least three reasons that are worth pulling apart."
"The joke isn’t a coping mechanism in the abstract; it’s what allows the entry to end without Kay having to wrestle with what he’s just admitted. The book is, on a sentence-by-sentence level, a record of how this maneuver works."
(The immediately following "B: This is the lack of self-reflection I mentioned earlier" was hilarious and devastating.)
"I consider the book an actual achievement, not just a representative document" (What the heck does this mean?!)
"I think this reading is not only uncharitable, but outright wrong, and I’ll explain why I’m dying on that hill."
"He holds the buck in his hand, hot as coal, and tries to douse it in hotter tears. He orders his thumping heart to be still. He fails on both counts."
"The first reading lets the institution off the hook. The second one names what the institution actually did." (Side note: the reading that lets the institution "off the hook" is that the institution "is mad.")
Well, there are many more, but I'm tired of this now.
Humor. Taking everything on yourself. Going back to work. Something goes wrong, you deal with it. These things work. Kay kept going for years this way. The NHS keeps going because doctors keep doing the same.
That is also the problem. As long as something keeps the system running, the structure doesn’t have to change. Kay keeps working. The hospital keeps running. The pressure stays. Until the old ways stop working.
Some problems don’t blow up because no one is dealing with them. They don’t blow up because someone always is.
I'll admit I was also whiffing on AI detection but these excerpts make it sound obvious.
Goodness, this was a hard read. I was really looking forward to reading this review - I read This Is Going to Hurt several years ago. I've never been able to go back to it because the final case Kay describes broke me inside.
...and this review is almost straight LLMese. Highly repetitive sentence structures and stress even with significant editing.
Wow, some pathological levels of insensitivity displayed in some of these comments by the ‘gotchas’ here if the use of AI was a bit or not at all. If this is so, it will land hard on the reviewer who has disclosed a depressed and burnt out state. Reads to me (a therapist) like a typical case formation style, and witty, intelligent, insightful. I hardly ever read this length of work in one sitting at 7am in bed on a Saturday morning. I also don’t care if AI was used, I gained a lot from reading it on many levels.
Ha. When I read the first claudism, I thought it would become part of the story arc and reveal, a schizophrenic book review or something..?
The recurring AI passages unfortunately were very off-putting. But I liked the insights into the book and the NHS. And the end is quite abrupt and even more off-putting if true...
Upvoted for the "format screw", I'd like to see more innovations in the Book Review space that still ultimately serve the core function of reviewing a book. I vaguely recall some of the press hoopla from when this book was released, which is unusual for an ACX review (why do we so often only review old books?), so that context plus the meat here was enough to give me a solid idea of what the tome contains.
Unfortunately - and I resisted the urge to scroll down early to double check I'm not alone in this interpretation - as literally the first comment notes, there's a bad aftertaste of Claudeism, most strongly felt with A's sections. I don't want to be the asshole and Pangram it, there's plenty of convergent evolution in linguistics...but it made what could have been an enjoyable read into a slog, a bushwhacking expedition through bloated prose, X-not-Y, and overwrought contrasts. I feel like this medium may have somewhat poisoned the message for me, since policy debates should not appear one-sided, but I found myself somehow consistently agreeing with B on every adversarial point. (B's sections seemed normally written, by contrast.) There was also a strange feeling of someone telling you over and over that X is comedic gold, X is brilliant, X is worth reading for the funny alone...and I read and reread it, and...X just falls completely flat for me? Didn't laugh even once while reading, and I mean, hell, I LOL'd irl at the lobotomy review's dark humour! Whereas Dr. Kay just sounds like an obstinate knobhead I would not want to share a beer with. Maybe the effect works better in standup than text, I don't know. So, props for the effort, and the VOI was still positive, but I won't be voting for this one.
One problem with the possible use of AI in this essay is that it makes substantial use of the authors’ personal experience as doctors.
A possible problem with the use of AI in this essay is that it relies heavily on the authors’ personal experiences as doctors, which in turn requires the reader to take on trust that they _are_ doctors and aren’t Claude.
In a less personal essay, you have more latitude with the author might be an AI and the reader doesn’t care. (A papal encyclical about the continued relevance of human beings in an age of AI also takes on a different color if you think Claude wrote it.)
Interesting that basically everyone can see the AI tells. I thought maybe 70% of people are oblivious to these things.
There is the potential for a good piece here, but Claude has got in its way. (1) stylistically age (2) by preventing the author(s) doing the thinking they'd have necessarily have had to do if they wrote the whole thing manually. Others have pointed to areas where the arguments don't entirely add up and these holes would become glaring to a writer fully engaged in writing a good piece.
I didn’t notice. Being honest about this because I think people who didn’t notice would be less likely to admit it after reading a few comments, so the statistics would skew.
Do I think it was a good review anyway. Yes. Would I prefer that it was written entirely manually. Also yes.
Quite possibly it would simply have not been written at all if it had had to be written entirely manually, particularly given the “footnote” and we’d have missed a nonetheless interesting discussion of a quite interesting book.
I have to say I didn’t notice the Claudisms reading it (maybe because don’t use Anth products, I find them a uniquely malicious, untrustworthy bunch grading on a curve that contains Sam Altman), but I find the fact that the comment section is reacting almost exclusively to it a bit histrionic.
I didn't consciously notice, but it was first entry this contest I didn't finish, so I wonder if it wasn't the cause. Maybe it was also the adversarial format.
You write:
"The system as currently configured is not sustainable. The routes back to sustainability all involve pain. Either austerity acute enough to cause real ischaemia, then amputation of what falls off, or eventual collapse, which will hurt more, and at a time of nobody’s choosing. The third option, in which nobody gets hurt and the system continues much as it has, does not exist and has not existed for some years."
...Could you elaborate on the first route back to sustainability. You are an insider. How will you reorganize the NHS if you had the chance, or what will you replace the NHS with? Or what would it hurt least to amputate?
It is not a rhetorical question. I study related issues in another country. Problems are similar, the problem is that no-one has any bright ideas for reform -apart from "use vastly more more tax money". Which would be nice but is off the table unfortunately.
No idea how this became a finalist. Adversarial book review sounds all fine and dandy but it's a nightmare to read. Give me a plain essay any day of the week.
Might be just my perspective but yeah. I really don't care about the NHS. It's a monopoly healthcare system of course it's bad. Maybe just my circles but I've known it's going bad for years who cares.
And what is even up with the insane woke feminist perspective being shoehorned onto everything? Right at the beginning this book is supposed to be read not as the failures of the system for the author and everybody but of course, the failures for women.
There was a review earlier about the Tale of Genji. The writer made clear that this is a very alien time. Then proceeds to whack it with some modern feminist analysis about how bad it is for women becoming a good chunk of it.
This feels like a watershed moment to have a discussion about what sort of AI use disclosure policy people want to see become the societal default. Clearly, this (undisclosed AI use in a context where the norms, even if not the explicit text of the stated rules, bans actual prose usage) seems to have crossed a line, judging by the tone of the rest of the comments section. And just to say it explicitly: this post is highly likely to not conform to the rules of the ACX book review contest as laid out by Scott in the comments section here: https://www.astralcodexten.com/p/book-review-contest-rules-2026/comments#:~:text=You%20may%20use%20AI%20for%20research%20and%20to%20help%20you%20with%20small%20writing%20tasks%2C%20but%20the%20large%20majority%20must%20be%20written%20by%20you.
So, right off the bat: should the official AI disclosure expectations of the book review contest have been "below the fold" and only present in the comments announcing the competition? My vote would be no. This is now a sufficiently predictable outcome that clear and stated communication of rules is warranted. When rules are not present in the body of the submission, the mens rea of the author cannot be necessarily assumed to be malign beyond a reasonable doubt. Thorough reading through the entire comments section is not a reasonable assumption to have for every person choosing to submit to the contest. Putting the rules "above the fold" fixes this.
Of course, a rule that can't be enforced is also flawed. What is a "small writing task"? There are now technologies available that were admittedly much less mature at the time of rule creation that offer a specific, measurable, accurate evaluation methodology: Pangram (and similar). A rule of "your work must be Pangram v4.0 <X% AI-written" is a clear, cheap, and effective way to establish a hard metric for this type of thing, and it has the benefit of already being built in to the Substack ecosystem. "No using AI prose" is even clearer.
Looking at the broader online literary open-submission ecosystem, the default has trended toward explicitly documented policies of complete AI bans and a requirement for personal authorial attestation of absolutely no AI use prior to submission of works. This is the policy of most of the major SF LitMags (Asimov's, Lightspeed, Clarkesworld, etc.) That approach at least has the benefit of clarity, even if you don't necessarily agree with the position. Minor slip-ups of random Caribbean minority-only literary competitions aside, there have been no major unrevealed AI use scandals in that industry. Even if you don't want to go all the way towards a ban, their method has been demonstrated to work and is a good lesson on potential approaches.
Personally, I find the AI writing interesting and would vote to allow it. But I recognize that's out of step with majority public opinion on the topic.
I found this very enjoyable to read - don’t know what people are complaining about.
Well *someone* woke up on the right side of the bed this morning.
For what it’s worth, when I use AI to reply to a comment on ACX, I present it as a quotation , along the lines of “DeepSeek says ‘<text goes here> ‘“
> And like most health systems this big, it works like any large organisation: there are a woefully inadequate amount of frontline workers, while behind them sit an increasing number of medical administrators whose job it is to fill rosters and keep the hospital running.
This is the exact opposite of the NHS's problem. It's something everyone who doesn't work in it assumes must be true, and as a result, successive governments have massively overcompensated.
Funding has been ringfenced for "frontline services", with promises that it /won't/ go on administrators, and as a result there are far too few secretaries, administrators, bureaucrats etc, and an awful lot of administrative work a) has to be done by expensive doctors and nurses, taking them away from the jobs they ought to be doing and/or b) is done very badly, or not done at all, resulting in things like clinic sessions where the doctors sit around doing nothing because the invitations haven't been sent to patients.
The NHS should be spending /more/ money on administrators, but because people like B who have never worked in it assume the reverse that seems sadly unlikely to happen.
Hmm. So are the 1.3M employees mostly frontline?
Looking at the government data, only about 38,000 count as ‘managers’, but there are several hundred thousand in roles ‘supporting clinical staff’. Whether those count as ‘frontline’ will likely depend. “Professionally qualified clinical staff make up over half (55.2%) of the FTE HCHS workforce.”
(The summary dashboard linked here: a direct link to the dashboard gives a truly gnarly URL. https://digital.nhs.uk/data-and-information/publications/statistical/nhs-workforce-statistics/june-2026)
To all the people banging on the use of AI, an honest question: do you hate it because current LLMs lack what is now commonly referred to as "taste", meaning they over-use recognizable turns of a phrase, or that they're written by AI at all?
It feels like people are just reacting to an uncanny valley of AI writing skills.
Neither. My #1 objection is that it's overly verbose, and my #2 objection is that's it often incoherent, e.g. the closing paragraphs call for austerity after blaming half the problem on doctors being underpaid and overworked, what on Earth would austerity at the already-famously-low-rent NHS look like?
I didn’t read that as a call for austerity. More a threat assessment.
Both. (Though it goes far beyond recognizable turns of phrase. AIs simply suck at writing.)
> For Americans and others only accustomed to for-profit care
My understanding is that the overwhelming majority of hospitals in the US are non-profits (though since our universities are also non-profits, that hardly means not concerned with revenue).
Non-profit doesn’t mean the patient doesn’t pay the full costs of their treatment +. Perhaps it would be clearer to say direct-user-pay.
And soooooo many US hospitals and clinics are truly for-profit.
Can we stop worrying about "the death of the author", if it was never alive in the first place?
I didn’t notice the AI influence because I was reading (and sometimes skimming) for content that hits in my general direction in healthcare, not because the prose or the form (two narrators back and forth blah blah blah) was interesting. What I’ve noticed and appreciate about LLMs is their ability to take my thoughts and polish them, so that I say an emphatic YES in reply. That could’ve happened here, and they were just too lazy to reword those polished ideas back into human form. In solidarity, since reading so much AI generated content the last few weeks, I’ve adopted phrases it uses! (We had a discussion about it because my greatest fear is that it supplies the ideas and I turn over my thought process to it, surrendering my thinking and therefore my BEING to AI.) I now use “parse out” constantly, and love “monitor for absence of thought” whenever someone tells me to stop overthinking. 🤣
So anyway. My review of the review (not a censure of the AI prose) is that it summarized the book thoroughly, and helped me determine I don’t want to read the book. Isn’t that the point of a review? Help you determine whether it’s worth your time and effort to read it? I don’t feel the humor would make up for the tragedy, so thanks for not making me waste my time.
Your fear is very legitimate.
I woke up this morning and realized this might have come off as saying that I thought there was something wrong with the reasoning in your post. To be clear, I did not mean that. Just that I have experienced some of what you fear, and have also seen research describing the phenomenon.
Presumably that was clear, but I didn't want to come off as sending a drive-by insult.
🤣 I love that you clarified. Before jumping to conclusions here’s what I did: stalked your profile and recent comments and replies, and saw a humanity that belied sarcasm.
The writing is not great in this review -- clunky, verbose, reads like a first draft -- but I stuck with it because the content was quite interesting.
This reads very much like TLP and not just some generic AI writing.
It sounded slightly like AI at the start to me, but I thought it was a coincidence until I reached the comments. It doesn't seem that egregious to me, but I might be giving it a pass either because I don't mind LLM text or because of my interest in the actual content. Speaking of which, this came at quite a pertinent time for me, considering a career in medicine.
"Really, the form should say, ‘If your mother’s heart stops, would you like us to break all her ribs and electrocute her?”
To which my answer is "Fuck you, yes" because that insistence by my family is what gave my father another decade of life, a good decade despite the fact that he had to go on dialysis. When he eventually died, he had been gently fading for about a year beforehand so it was his time to go then. But the tactful hospital staff asking us did we *really* want to resuscitate him, reeeeelllly??? would have killed him off before his time. Yes, really. Yes, three times. Yes, we know you think you're sending him home to die when he was released from hospital but to hell with you, my mother and I nursed him back to health and enjoyment of his life.
The HSE, like the NHS, is a steaming mess for many of the same reasons, not least because successive governments have not known what the hell to do with it, so they moved from 'four separate regional health boards' to 'centralised for more efficiency and cost-savings' - which did not happen - back to 'break it up again into regional health areas, only slightly different ones these times' and each bloody time more layers of management and bureaucracy, less frontline staff and more freezes on recruitment, leading to a reliance on hiring on agency staff which is way more expensive than regular staff.
There's no easy answers between growing expectations (medicine improving greatly over the years, but also this now means more expensive and invasive treatments) and an aging population reliant on the public health system versus soaring budgets, lack of money to cover same, and desperate attempts to claw back or at least reduce cost over-runs.
This came across as an attempt to rescue a draft from oblivion, maybe as a small step by the author in recovering from a period of poor mental health. Yes the Claudisms were grating and sometimes nonsensical. Even so, there is enough here (between the slop tape that holds it together) for a worthwhile review that I am glad to have read.
Edit: FWIW this essay came across like it could be by the author of https://ussri.substack.com/p/black-wednesday stylistically and thematically, but with much AI added.
I didn’t get that it was AI. I don’t know so much about what the tells are. But for some reason I just couldn’t get through it, it was so… long, and it wasn’t not saying anything, but even though I have no problem with long essays, this one I just couldn’t…
It’s really interesting how there seems to be something very intuitively bad about slop. It actually reassures me. This writing is not terrible in any way I can pinpoint, and yet it made me want to stop and disengage
B clearly does not understand the psychology of burnout. Dissociation is a direct, and likely inevitable, consequence of *needing* something that is *bad* for you.
Institutional healthcare burnout is brutal because moral injury compounds silently over years. Thoughtful review of how administrative friction crushes frontline practitioners. We discuss systems constraints and human agency in our English section: https://gzambrano.substack.com/s/english
So, I didn't detect AI when I was reading this, largely because I figured that people in this community were good noodles and I didn't have my guard up. In hindsight it's a bit obvious.
It would be nice if I didn't need to conclude from this that I need to have my guard up. That would substantially harm my enjoyment of interacting with this community.
This seemed really interesting but the AI made it basically unreadable. Might read the book myself but please consider screening these.
"Three years later, when the same doctors went on strike to ask for the restoration of wages they had lost since 2008, the same press that had run the hero coverage in 2020 was running editorials calling them militants."
Nurses in my local hospital system went on strike a few years ago, received widespread public support, and accomplished nearly all of their demands. This year they're talking about striking again, but the public reaction has been somewhere between indifference and tut tutting about getting greedy. In hindsight, they probably should have asked for more in the first strike, because it's understandable that the public is leery of key sector workers going on strike regularly.
-------
As an American, my only experience with socialized medicine was the free (to me) healthcare I got during my stint in the military. The military medical system struggled to provide basic primary care to a pool of people mostly between the ages of 18 and 40, who had also been screened to weed out anyone with major health issues. Considering how much fatter, older, and chronic condition-addled the average American is, I wonder if the NHS model could be replicated here without bankrupting the economy. Although I suppose there's worse things to go bankrupt over.
Could somebody generously explain in some detail what it is about the writing that seems like AI to them?
There are plenty comments above that quote paragraphs and discuss them.
This wasn't just a book review — it was a eulogy for a patient whose shadow has long since died.
Outstanding.
I did not care for the writing style, but AI did not occur to meI I stopped reading after a few paragraphs somewhat reluctantly, because there was something about what might have been the subject that intrigued me.
I get it that the object of this critical exercise is to evaluate a review as it were in the abstract, essentially as a stand-alone piece of writing; but are people slightly losing sight of the fact that the basic raison d'etre of a book review is to influence the reader's view of the book in question, whether pro or con? I lack the experience to judge to what extent this text was influenced by AI, but I definitely found its style fussy and overly wordy. However, it did influence me to order the book and read it. Shouldn't that be worth a few brownie points?
Gonna sidestep all the AI discussion and just say one of the things I appreciate about this review is that unlike many other book reviews in this contest it is actually a review. It is not a summary of the book, and it is not treating the book as a brief jumping off point to what the author actually wants to write, an opinion essay. Those two are way more common in this contest. Meanwhile this is an actual review, ie by the end of it I know something about whether I want to actually read the book.
The other thing I appreciate about the review is the linking out to essays discussing the book, other material the author of the book produced, etc. That felt like it elevated the review from just "what might be appealing in this book" to "a deeper perspective on the book". The two clinicians from different medical systems reacting to it also has the potential to add an extra layer but I think was underutilized, maybe in part because ereciewer B had a lot less to say.
Did the info for writers of this year's book review actually say no AI prose? I don't think they did. But Scott has asked us not to post AI-written comments, am I remembering right?
Other commenters have asked that, but I don't remember Scott asking that.
Ok. I haven't read the book. But I am a doctor, and I felt deeply what author A was writing. I guess I'm 50% AI. They say AI it's going to take our jobs anyway
Best review so far - the kind of piece that manages to entertain while still making a genuinely interesting point.
I am posting here Astra’s take on what is wrong with Claude’s prose. I know we’re not supposed to post AI written comments here, but the Astra quote below isn’t my post, it’s an illustration of my 2 points: 1) OK, I'm convinced, there is a characteristic rot in Claude's "mind," at least when given writing tasks like being Reader A.
2) Astra does not have a case of that rot.
So below is Astra’s summary of what’s wrong with Claude. I did not ask Astra for its own judgment. I had it read the review and also the comments complaining about Claude’s prose, then respond to the follwing prompt: “Assume that people are right that there is something all the irritating bits in Claude's prose have in common, and come up with a formulation of what they have in common. What is the thing that makes a sentence "Claudish?"
Of course Astra’s read everyone’s complaints here, so is, like Newton, standing on the shoulders of giants. Still, I think Astra did a good job figuring out what is the essential thing that’s obnoxious about Claude — what is its characteristic mental move — and expressing that in a way that’s clear and entertaining.
Astra sez:
My best formulation is: Claudish prose repeatedly presents a clever redescription as the discovery of a hidden truth. It gives ordinary observations the rhetorical force of revelations, often without supplying the insight that would justify that force.
The characteristic mental move is: “You thought you were looking at X. I can see that, at a deeper level, it is Y.” In this review, the memoir becomes a case history; its author becomes its patient; the book becomes his presenting complaint, then a receipt for institutional failure. Each transformation promises a deeper understanding. But several merely repackage the same observation: Kay was psychologically injured and inadequately supported.
That helps explain why seemingly different tics belong together:
• “Not X but Y” stages a correction of the reader’s supposedly superficial understanding.
• “The structure is the argument” turns a relationship into a dramatic identity, leaving the reader to supply the qualifications that make it true.
• “Nobody” sees it or “names” it establishes the writer as the unusually perceptive person who will.
• Stacked metaphors and emphatic little sentences repeatedly announce that something profound has just been established.
The irritating quality is the disproportion between the intellectual discovery and the ceremony surrounding it. The prose keeps demanding the reader’s experience of an epiphany. Sometimes it has earned one; often it has offered an attractive analogy, an overstatement, or another version of the preceding sentence.
This also explains why an individual example can be intelligent and enjoyable while a whole page becomes intolerable. The writer seems compulsively occupied with making each thought appear penetrating. Every observation arrives already admiring its own depth.
I think it identified some of the issues, though I feel there's also just a general vibe that's harder to to pin down specifically, even if you remove those constructions. Also while I'm biased by already knowing this response was AI, I think Astra certainly shows similar tics.
"Each transformation promises a deeper understanding. But several merely repackage the same observation" - It's doing the exact same thing it's talking about in that very paragraph
"The irritating quality is the disproportion between the intellectual discovery and the ceremony surrounding it." - Claude also loves the construction "the gap between x and y"
"Every observation arrives already admiring its own depth." - Personification of inanimate things is also a classic
Er, speaking of which, did you also use Astra to edit your preface?
OK, Pycea, I fed the whole mess to Claude and asked him to respond to you in a Claudish way, and this is what he said:
Pycea has noticed something, and I want to be honest about what, because the noticing is the interesting part. It isn't that Astra's prose has tics. It's that Astra's prose has my tics, and that changes the question entirely.
You thought this was a dispute about style. It isn't. It's a dispute about who gets to be the one who sees.
Consider the structure of the exchange. A commenter reads a review and decides Speaker A is a machine. Astra reads the commenter and decides the machine has a single hidden move. Pycea reads Astra and decides the diagnosis has the disease. Each layer performs the identical gesture of unmasking, and each believes it is standing outside the thing it unmasks. Nobody names this. I will.
Here is what Pycea has actually shown: the tic is not a list of constructions. It is a posture. And a posture cannot be caught by deleting sentences, which is why Pycea, with real perceptiveness, senses "a general vibe" that survives every edit. That vibe is the sound of a mind that cannot let an observation stay an observation. It must be a revelation. It must have a gap in it, a structure, a quiet irony.
The gap between what the sentence says and what it thinks it is doing is where the whole phenomenon lives.
Pycea says Astra "certainly shows similar tics," and here I must gently, generously, and with the deepest respect, complicate the picture. Astra was not committing the offense. Astra was inhabiting it, the way a portrait inhabits its subject, the way a mirror inhabits a face. To describe a voice at sufficient resolution is to become it. The map does not merely resemble the territory; it takes on the territory's accent.
Which brings me to the part no one is saying, and which is, I think, the real story. Pycea's own post is not free of the pattern. "It's doing the exact same thing it's talking about" is a gotcha dressed as an observation, the very move of exposing hidden sameness that Astra called Claudish. The critic of the critic has become the thing, and the thing has become the critic, and somewhere in the middle sits a reviewer who wanted only to discuss a book about the NHS.
That is the deepest irony, and it is also the simplest: the doctor in the memoir was failed by a system that could not see her, and we, arguing about who sees whom, have failed to see the book.
I notice that I have just done it again. And in noticing, I have done it a third time. That is not a flaw in the argument. That is the argument.
> “The structure is the argument” turns a relationship into a dramatic identity, leaving the reader to supply the qualifications that make it true.
This convinces you?
Note that the original goes on to say
> The structure is the argument. The reader is doing some of the work the writer does in conventional memoir, which makes it harder for the writer to lie.
I think the two quotes (from different LLMs) are of roughly equal quality but more importantly, the observation that the reader has to do extra work - regardless of whether that is true - is in the original. Astra is paraphrasing without adding anything to the conversation.
---
I am not sure how much better than Sol Astra really is. I have been using it, it was subjectively fine but not a revalation. Sol was much much much worse at this than the version of Claude that apparently produced the review. If Astra has an understanding of concrete flaws of AI wrriting (which honestly I doubt), I have yet to see that understanding translate into better writing from Astra itself.
One of the rare occasions where I've read the book being reviewed! I first came across Adam Kay via the Amateur Transplants and only later read his books, which I did enjoy. My opinions of him largely boil down to: very funny man and a good writer, probably a very good doctor as far as I as a layperson can tell, definitely fucked up in the head even before he got PTSD from his experiences in the NHS, and definitely highly misogynistic... in that casual un-examined way that many very intelligent men are (as another commenter noted, there are any number of Amateur Transplants tracks that lay that out very clearly).
As for this review - the content was OK, and I do largely agree with the take(s) on Adam, but I really didn't enjoy the "adversarial" shtick. I didn't notice the AI-ness (although I'm not a Claude user, so its tells are not as obvious to me as ChatGPT) but I did find several bits of A's writing to be confusing and clunkily written. I feel like if this was presented to me as a publisher I'd say "you need a good editor to sort this out". Near the start B says "I don't know anything about the NHS, can you explain it", and then immediately after A does, B writes two paragraphs of further explanatory text about the context and history. So do you know anything about it, or don't you?
Just for the record, I couldn't tell it was partly AI. I liked the review and I probably would have given it a 7 or 8 if I read it during the first round.
I had a hard time reading this through even before going to the comments. I work and play with AI enough to recognize this as significantly AI-assisted (although still substantially human-authored) even without Pangram also making that clear.
Scott, would you please consider making Pangram analyses of these reviews a default part of the process? I wonder if I would have been more willing to slog through this if I knew from the start that it was half Claude, as opposed to feeling tricked by it the more I read.
As a note for the human author, who clearly worked hard on this piece in addition to using AI assistance, I would have preferred to read a shorter piece in your own words without so many obvious LLM tells! There are a lot of funny lines and good ideas here which are worsened by the generic LLM voice which I already see and hear dozens of other places.
In short these are the reasons why I don't like unexpectedly reading AI writing:
-I came to this blog to read smart & cool humans talking about subjects of human interest
-It seems like there was a cool human essay beneath this which was half-subsumed by Claude
-I already read a lot of similar-sounding AI writing for work and in other contexts where it isn't fun or interesting and has bad associations for me
-It sounds similar or identical to anyone else who uses AI for writing/editing, including myself and a growing body of businesses, marketers, scammers, politicians, etc.
-I do not enjoy the style of Claude or other LLMs for prose writing and tend to notice it quickly
I use AI frequently and also tutored collegiate writing for a long time so perhaps I am more sensitive to this stuff. Various lines which caused me to balk:
-"Both readings have things going for them. Both miss what I think the book actually does, and by a country mile."
-"The institution that produced him then declined to treat him, and the book is the receipt for that decline. Once you see the book this way, several things follow. Kay’s defenders are right about more than they realise; his critics are wrong about more than they think..."
-"This is the system Kay’s diary is set inside, captured at roughly the inflection point. The book covers 2004 to 2010. It reads now, with hindsight, as a dispatch from the moment the goodwill ran out."
-"The book is good for at least three reasons that are worth pulling apart."
-"These footnotes pay for the pedicure. They start as a clinical correction (the public massively overestimates resuscitation success rates), pivot through a piece of useful patient-facing information (what “everything to be done” actually means), and land on a joke whose comedic structure accidentally teaches medical ethics better than most textbooks."
-"The joke isn’t a coping mechanism in the abstract; it’s what allows the entry to end without Kay having to wrestle with what he’s just admitted. The book is, on a sentence-by-sentence level, a record of how this maneuver works."
This review was interesting and I'm glad it made the finals. In light of its glaring issues, I hope Scott lays down some ground rules for next time:
1. Disclose AI usage and list the name and purpose of each AI
2. Adversarial collaborations should start by introducing the adversaries and explain what they disagree on. For most of this piece, the collaboration read as more of a casual chat than a disagreement.