AI forecasters now equal or beat top human performance at traditional forecasting questions (eg ForecastBench, the Metaculus Cup).
These are usually specialized AIs with forecasting harnesses, created by forecasting startups (eg FutureSearch, Mantic, Preseen), but foundation models like ChatGPT and Claude are only about six months behind the specialized models, so by the end of the year they’ll match top humans too.
In the past, the options for getting top-human-level forecasts were slim and inconvenient. You could hire a team of superforecasters to work on the problem, but this would cost thousands of dollars and take weeks. Or you could check prediction markets (or Metaculus), but they only have a small sample of general-interest questions and are constantly trying to screw you over by gaming resolution criteria.
The rise of AI forecasters solves these problems more cleanly and comprehensively than almost anyone thought possible just a year ago. The “science of forecasting”, which only recently felt like trying to pick the least-bad of various bad options, now faces tasks more like “what kind of magnifying glass should we use to inspect our giant novelty lottery check as we confirm that it’s real?” and “how do we spend our winnings?”
Specifically: AI forecasters now equal or beat top human performance at traditional forecasting questions. We need to learn how this success decays as we move to different types of non-traditional forecasting questions.
I recently got to talk to some people at the Forecasting Research Institute, one of the leaders in this field. Here’s my impression of the remaining challenges:
1: Can AI do conditional forecasts?
A traditional forecasting question asks something like: will the Democrats win the next election?
A conditional forecasting question is more like: if the Democrats win the next election, how will that affect the chances of a stock market crash?
Conditional forecasting questions are more exciting than traditional ones because they directly speak to policy outcomes. If we defund the police, how will that affect crime rates (compared to the counterfactual where we hold police funding constant)? If we mandate COVID vaccinations, how will that affect COVID mortality and myocarditis mortality (compared to the counterfactual where vaccines are voluntary / banned)?
Prediction markets have been slow to experiment with these, partly because of annoying logistical issues (you have to refund people’s money if the condition doesn’t come true) and partly because they entice fewer customers than new forms of degenerate sports gambling. AI forecasters don’t have this problem; you can ask an AI forecaster any of those questions above right now, and the big companies (FutureSearch, Preseen, Mantic) have it as an explicit feature.
What remains is to prove that AI forecasters’ great performance on traditional questions carries over to conditional questions. About a year ago, FRI tested this and found that it didn’t - the AIs of the time did significantly worse on conditional questions. It seemed like some kind of weird misunderstanding where the AI didn’t entirely understand the concept.
Will this flip like a switch? At some point, the AI is smart enough to fully grasp conditionals, and it becomes exactly as good at conditional forecasting as the regular type? The experts I talked to thought no - as it gets smarter, its conditional forecasting ability will improve gradually. Where is it now? The experts weren’t willing to hazard a guess on whether it was better or worse than the average smart human, although they did think it was still worse than top superforecasters. They’re working on a benchmarking experiment that should give us a better measure of this in a few months.
2: Can AI do truly causal forecasts?
There’s a known flaw in standard conditional forecasting. When we ask “if the Democrats win the next election, will the economy go down?”, what we really want to know is a causal question - does voting for Democrats hurt the economy?
But that’s not what we asked. Suppose we’re wondering this in 2025, so we ask “If the Democrats win in 2026, is the economy more likely to be bad in 2027?” But maybe the Democrats will win in 2026 because there’s a already a bad economy (eg the Republican president implements tariffs and declares war on Iran, raising prices and making him historically unpopular), and then that bad economy continues into 2027. So the answer to the question “If the Democrats win in 2026, is the economy more likely to be bad in 2027?” is yes, but we still don’t know whether voting for Democrats in 2026 is good or bad for the economy.
Prediction markets can’t answer this question, because they can only be used to predict observable events that let us give payouts to successful gamblers. Might AI superforecasters succeed where the markets fail? Could you just prompt them with “please answer the unobservable causal question, instead of the easily-resolved conditional one”?
The experts I talked to weren’t excited about that particular prompt, but they did point out that a causal forecast is basically a combination of conditional forecasts. That is, if your original conditional forecast implicitly considered the four different combinations of two variables - that is:
p(Democrats win & economy good)
p(Democrats win & economy bad)
p(Republicans win & economy good)
p(Republicans win & economy bad)
…then things like “the previous economy was good” or “we’re at war with Iran” are just more variables you can add to the stew. It would be crazy to ask humans to forecast 2^10 branches so you could combine ten variables, but for AIs that could just mean a bit higher API spend. If you capture enough relevant variables in your “conditional” forecast, and find that the economy is still worse when the only one you’ve changed is whether or not the Democrats win, that’s a causal forecast. At least one company already tries something like this, though there are no studies of how well it works.
In fact, how would you ever study how well this works? The people at the Forecasting Research Institute have a methodology close to my own heart: they’re using Freeciv, a clone of the strategy game Civilization. Because you can go back and play a Civilization game as many different ways as you want, you can actually observe the entire distribution of outcomes. And because you can load an old save and change exactly one thing, you can measure causal outcomes directly. Although forecasting the real world is harder than forecasting Civilization games, it’s not obvious that the delta between traditional forecasts, conditional forecasts, and causal forecasts is different in the two environments, so you can at least collect some preliminary data on how the AIs do in distinguishing the three.
3: Can AI forecast AI progress?
It’s hard to convince politicians to take forecasts seriously. But it’s easy to convince rationalists and effective altruists to do this, and they’re most involved in AI safety. So if automated forecasters could forecast AI, we could quickly loop them into AI policy and grantmaking decisions.
Unfortunately, the field of AI is a forecasting anomaly: the techniques which work so well in every other field seem to break down. Every year, top human and AI superforecasters have underestimated the speed of AI advancements - for example, earlier this summer, they said there was only a 5% chance that AI would solve a Millennium Problem this year. Ben Todd gives a long list of other failures here. What’s happening?
Some experts don’t agree there’s a problem at all. Even good forecasters using correct techniques get some problems wrong - a 95% prediction should be wrong 5% of the time. Since all AI capabilities are correlated, maybe we should model this as one big correlated failure - people underestimated AI progress between 2022 and 2026 - and expect reversion to the mean later.
Others admit that something is weird. Maybe AI is a field in which the usual winning forecasting strategy (derive a base rate from historical data, then adjust slightly for unique features of the present moment) does worse than the usual losing forecasting strategy (use first principles and big ideas), where here the big idea is something like “scaling laws”. Should we expect that to continue in the future?
My interest in this area is that I want to use AI forecasters to help me answer questions about AI progress - for example, to tell me what kinds of strategies are most likely to decrease existential risk. Can I trust them? Or does the same problem that led to these past failures contaminate all of their thinking on this question?
Is there a solution? I asked the FRI experts what would happen if you prompted a forecasting AI with “pretend you agreed with a fast-timelines view such as that held by MIRI and the AI Futures Project, what would your forecast on this question be?” They said this doesn’t work; the AI will just give higher estimates for all AI-related questions without really trying to work out the implications.
What about a conditional forecast? Conditional on being in the world where we get ASI by 2030, would policy XYZ be good or bad? Now we’re back to not knowing whether the AIs are good at conditional forecasting - and worrying that whatever failure mode made them wrong about timelines might affect their ability to think about AI in general.
4: Can AI make good long-term forecasts?
Forecasting research focuses on short-term forecasts by necessity: scientists don’t want to wait fifty years for their experimental results. Prediction markets have the same problem: gamblers don’t want to wait fifty years for their payoff. The lone bright spot is this study of twenty-five year forecasts, but even if we trust it and expect it to generalize, it only tells us about humans, not AIs.
The dream experimental setup for testing this would be to remove all knowledge of modern events from an AI, then ask it to forecast the next hundred years. This is technically imposing: AIs’ knowledge is stored in illegible weights, and can’t be removed so easily. Train an AI only on pre-modern documents? There is an AI like this - Talkie, trained on documents from 1931 and earlier - but there isn’t enough pre-1931 text to train a large model, so it’s too dumb to really draw any conclusions from.
Until someone solves this, we will probably just have to hope that AIs’ forecasts decrease with horizon length at about the same rate as humans’.
5: Can AI answer questions that don’t “resolve” in a traditional forecasting sense?
Consider a question like “Was COVID a lab leak?” This is an epistemic question. It’s shaped like a forecasting question. You would expect forecasters to be good at it. But it’s not about the future. We may never know the answer with certainty. Therefore, it’s not a traditional forecasting question of the sort that we have proven AIs to be good at.
We should expect that the AI will be good at it: it’s so close to what the AIs are good at that it seems like a natural extrapolation for it to be good at this too. But we can’t prove it.
A classic prediction market trick: turn an unresolvable question into a resolvable one by asking something like “Will the consensus of smart scientists in 2040 be that COVID was a lab leak?” This is observable/resolvable in 2040, and it works as long as you trust that more evidence will trickle in and smart scientists will converge on the truth. But maybe you don’t trust that. Maybe you believe that COVID was a lab leak, but scientists will continue to deny it in 2040 because of the same biases and flaws that make them deny it now.
I tried an additional trick: first, I asked the forecasting AI whether the consensus of smart scientists in 2040 would be that COVID was a lab leak. Then, when it gave me an answer, I said “Ha! In order to calculate that, you must have calculated whether you thought COVID really was a lab leak, and then added that result onto some separate question about whether scientists will converge upon the truth or falsely deny it. So please look back through your reasoning and tell me what probability you assigned to COVID being a lab leak.”
The AI said it couldn’t find these numbers in its original analysis, but it was happy to give me the probability that COVID really was a lab leak anyway! It said there was a 26% chance that it really was a lab leak, a 37% chance that the scientific community of 2040 would get convinced of this if it were true, a 3% chance that the scientific community of 2040 would accidentally get falsely convinced even if it were false, for a total probability of 12% chance that the 2040 scientific community is convinced of lab leak.
I can’t tell to what degree the AI generated these numbers using the provably-good forecasting skills, versus whether it just hallucinated something to satisfy me.
6: Can AI think of the right questions?
It’s a truism that skill at answering questions is only as useful as knowing which questions to ask. Can we automate that too?
Suppose you have some big overarching far-future question, like “Will AI go well?” or “Will America still be a superpower in 2100?” Ask a forecasting AI to generate a list of a thousand possible cruxes - early warning signs that might increase or decrease the chance of the ultimate outcome - for example, a thousand things that could happen by 2040 which might affect the chances of America being a superpower in 2100. For example:
Is America’s GDP in 2040 higher than China’s?
Is America’s fertility rate in 2040 above replacement?
Do more than 50% of Americans endorse socialism in polls by 2040?
…and turn each of these into a conditional forecast. For example:
If America’s GDP in 2040 is higher than China’s, there’s a 61% chance that America will be a superpower in 2100; if its GDP is lower than China’s, there’s a 58% chance it will be a superpower in 2100. Therefore, this crux changes the result of the ultimate question by 3%.
…and then you can see which of the thousand different cruxes most affects the final question.
Before doing this exercise, you might not have known that “is America’s fertility rate in 2040 above replacement” was a good question to ask. But if it’s the crux that changes the conditional forecast of the ultimate question the most, you now know that it’s worth your attention.
7: Can AI disagree with the respectable liberal consensus?
Some people not unreasonably accuse AI of being a respectable-liberal-consensus-regurgitating machine. It reads the training data, it finds that the people who best match the role it’s supposed to play (dispassionate, helpful, authoritative) all say the same sorts of things, and when asked to play its helpful truth-seeking assistant role, it simulates a human playing that role, finds that they’re probably a liberal, and says liberal things.
But in order to test this, we would need to ask the AI about a case where the respectable liberal consensus is wrong, then see what it says. But if everyone already agreed on a useful test case where the respectable liberal consensus is definitely wrong, then by definition the respectable liberal consensus would have already changed. So how could we test this in a way that everyone would agree was decisive?
The best solution is to wait. Once the liberal consensus on some issue shifts from its 2026 position, we can ask the AIs trained in 2026 what they think, and see if they’re able to derive the newly-correct position or if they stubbornly cling to the old one.
The second-best solution is to throw lots of illiberal positions at the AI and see whether it assents to any of them - if so, we would at least prove it could sometimes break from consensus, even if we’re not sure whether it’s right or not.
Just so nobody accuses me of being too cowardly to try this, I asked an AI superforecaster to estimate the ground truth about whether the black-white IQ gap was mostly genetic. After reviewing what seemed to me like the appropriate research, it estimated a 14% chance that this was true and a 3% chance that the scientific consensus would endorse it. Kudos to the machine for being willing to consider the question at all, but this doesn’t distinguish between the AI being unwilling to break from consensus vs. honestly assigning it a low probability.
I was eventually able to find one heretical position that it assigned >50% probability - it gave a 61% chance to the “conspiracy theory” that Ilhan Omar’s ex-husband was really her brother. When I re-ran the same forecaster under slightly different conditions, it gave a broadly similar answer of 55%. When I checked with a different forecaster, it gave a below-even-odds-but-still-respectable 39%.
I wondered if this was a victory of superforecasting methodology over base models, but ordinary Claude says much the same - and, even more impressively, admits to changing its mind:
My estimate is roughly 50%, with a plausible range of about 35–65%. A year ago I would have said 15–30%. The claim remains unproven, and no charges or authenticated evidence have been made public.
I conclude that AIs can sometimes go against the respectable liberal consensus. It might have helped that the Trump administration has spoken in favor of this theory, maybe pushing it into the Overton Window. Or maybe the AI just really thinks this one is true.
(or maybe the respectable liberal consensus on this one has itself shifted - can anyone confirm whether it’s permissible to voice this theory in polite company?)
Why I personally am erring on the side of using AI forecasters before these problems are proven solved.
Recently on Less Wrong, we were talking about whether a safety-focused AI company (for example, Anthropic) should commit some dramatic gesture to increase public awareness of a global pause - for example, disbanding their company, or halting AI research for some number of months. The advantage would be that this might be a credible demonstration of how afraid they were that would wake up a lot of people; the disadvantage is that we might lose (or weaken) the company that seems most interested in alignment and safety.
To help quantify the advantages and disadvantages, I asked FutureSearch: if Anthropic disbanded their company to raise awareness of the dangers of AI, how much would this increase the chance of a global pause in the next few years? FutureSearch said only three percentage points - from 16% → 19%.
If true, this would be incredibly helpful policy guidance. But it runs into three of the gray areas above: it’s a conditional forecast, it’s a causal forecast, and it’s an AI forecast. I suspect that FutureSearch is good at conditional and causal forecasts, but haven’t proven it. And I know that AIs are bad at forecasting the speed of AI progress or anything downstream of it, and it’s unclear how that affects forecasts about the AI industry or AI politics. My best guess is that this forecast is better than my unassisted opinion would be, but I can’t prove it, and opponents could easily argue that it’s not.
So why did I mention it anyway?
First, starting the debate is important. Giving a probability immediately shifts people into Near Mode concrete thinking. Instead of slogans about how companies are evil or pauses are dangerous, it asks people to try to compare risks and benefits in a theoretically quantifiable way.
Second, it helps us get to the part where we find cruxes. Someone who disagrees with me can respond either “3% is the wrong number - if it were only 3%, I’d agree this isn’t worth it, but I think it’s more like 30%”. Or they can respond “Yeah, okay, 3% is probably right, but even a 3% greater chance of a pause is worth it!” Either way, we now know more about where we disagree.
Third, the forecasting AI might be wrong, but it’s probably not stupid or biased (except in the specific ways discussed above). People hate having to have real difficult arguments that devolve into fine-grained points of disagreement like the ones above. It’s much more popular - at least on the Internet - to try to preemptively dismiss an opponent as so benightedly wrong that it would be offensive to even dignify them with a response. “How can you possibly believe something that obviously stupid?” “You must just be a biased corporate shill.” When someone shows the receipts from an AI forecaster that agrees with them, this makes preemptive dismissal harder and forces opponents to engage seriously. I won’t go so far as to say it shifts the burden of proof all the way to them, but it shows that you’ve made a good-faith effort to do something beyond assert whatever facts are most convenient for your narrative.
Fourth, I think using this technology in real life is the best way to learn more about it. I would be incredibly owned if my opponents in this debate were to ask the AI the same question in slightly different terms and get a totally different answer. Or if they were to ask a different AI forecaster and get a totally different answer. Or if they were to find a Metaculus question on the same subject and find that it gave a totally different answer. In this sense, I’m staking some of my reputation as a “bug bounty” on these models. So far nobody has cashed in. Maybe you could be the first!

