What Deontological Bars?
...
Constraint consequentialists believe that you should try to do good things that improve the world, unless those break hard-and-fast rules (“deontological bars”).
For example, you shouldn’t assassinate democratically-elected leaders, even very bad ones. Why not? Since bad leaders set bad policy, and bad policy can kill many thousands of people, wouldn’t it be for the greater good? Because there’s always one gun-owner who thinks any given leader’s policies are bad, so without the rule, every leader would face constant assassination attempts, probably some of them would succeed, and the nation would either crumble or degenerate into a security state.
This explanation combines two sub-explanations. In the first, you are wrong about whether assassinating the leader would produce good consequences - you think it would, but actually it would produce instability, tyranny, etc. In the second, you’re right - maybe you’re a brilliant forecaster who can see that this particular assassination would end with an orderly succession by a superior ruler. But you know that there are far more people who think they are such brilliant forecasters than who actually are, and you either use the Outside View to suspect that you are also deceiving yourself, or at least realize that the only stable bright-line equilibrium is for everyone - true brilliant forecasters and wannabes alike - to refuse to act upon their apparent foreknowledge.
“Don’t kill people” is a gimme. What other deontological bars constrain our actions?
I’ve been thinking about this lately because of an internal debate in the AI safety movement. Some people want to work with the least irresponsible AI labs, helping them “win” the “race” and hopefully do a better job creating superintelligence than their competitors. Others want to pause or ban AI research - the exact details vary from plan to plan, but assume they’ve already thought of and written hundred-page papers addressing your obvious objections. Different people have different opinions about which strategy is more likely to help, and it’s possible to coexist and pursue both at once. But in fact, both sides are a little nervous that the other is breaking a deontological bar.
Some of the people working on pause-AI regulations think there might be a deontologic bar against supporting AI companies. These companies are racing each other to create a potentially world-ending technology. If one company’s product has a 90% chance of ending the world, and another’s has an 80% chance of taking over the world, giving your money/support/encouragement to the 80%-ers seems kind of like endorsing evil. I don’t know if it was encouraged by this question exactly, but someone held a Twitter poll about whether you would become a concentration camp guard if you predicted you could get away with being only 90% as brutal as your average coworker. Taking the job would have good consequences, but is there a deontological bar in the way?
Some of the people working with the companies think there might be a deontologic bar against certain types of mass activism. The sorts of arguments that do well on LessWrong.com won’t give us landslide wins in national elections. That’s going to require things like working with Steve Bannon, working with Bernie Sanders, working with NIMBYs who hate data centers because they’re a thing that might be built in someone’s backyard (or by non-union labor), training TikTok influencers create short-form videos about the dangers of AI, and holding protests where we chant vapid slogans outside AI company headquarters. There are better and worse ways to do all these things, but once you lay out the welcome mat, you have limited control over who shows up - and every time someone tries to create the Peaceful Nonviolent Pause AI Movement Based On Peaceful Nonviolence For Peaceful Nonviolent People, it spends an inordinate amount of resources keeping out violent crazies who want to tag along.
I would be better positioned to navigate this debate if I knew what deontological bars were or where they came from.
A common formulation is “act as if your maxim would become a general law”. This correctly rederives the bar against assassination, but fails even in some simple cases. For example, should the Ukraine abolish its military? If this became a general law, it would be great - it would end all war, and countries could spend their military budget on domestic priorities (or lower taxes). But doing this unilaterally would be disastrous. But most moral decisions are unilateral, so what remains of acting as if your maxim would become a general law?
Trying to resolve this particular loophole gets us something like “Don’t be the first person to defect from a generally functioning norm”. If no other country had a military or was planning to create one, then unilaterally building a military (and forcing everyone else to catch up) would genuinely be immoral. This also covers our assassination example: in a world where assassination is rare and effectively punished when it occurs, making the attempt yourself is defecting from the existing norm. In some alternative world where one party constantly assassinated the other party’s leaders, maybe the other party would have the right to join in rather than be constantly defected against?
(one problem with this formulation is that it’s hard to operationalize “generally functioning norm”. There will always be too-eager people desperate for some sign that the norm has become “non-functioning” so they can violate it at will. For example, there have been several assassination attempts recently on President Trump. I don’t think this makes the norm against assassination “non-functioning” or licenses retaliatory assassinations by Republicans - I don’t even think a successful Trump assassination would do this - but I’m sure some paranoid person somewhere would have that thought. I don’t know how to get around this and think it might require common sense.)
But this still doesn’t do it for me. Consider: should I use my blog to spread misinformation if I think it would have good consequences? Maybe if I found a very effective charity I wanted you to donate to, I could play up the benefits and minimizie its downsides to make it seem even more attractive? Nobody can claim with a straight face that there’s a generally-functioning norm against this kind of mild online misinformation. But it still seems like a good theory of deontological bars should say that I shouldn’t do this.
The best I can come up with is “Don’t do something which would be bad if universalized, unless the norm is non-functioning in such a way that you’d be playing cooperate while your enemy plays defect”. This lets Ukraine avoid disbanding its military - they would be “cooperating” with a “norm” so non-functioning that their enemies would surely “defect” by keeping their militaries intact. But it doesn’t let me spread misinformation with my blog: other bloggers might be spreading misinformation, but this isn’t a defection against me in particular, and it doesn’t doom me to automatically lose if I don’t follow suit.
(although again, this requires some interpretive work. Suppose I learn I’m getting 10% fewer subscriptions than I would if I spread more information, or that my political priorities are getting 10% fewer votes/donations, and this is directly because other bloggers who spread more misinformation are competing against me and winning. What now? I think I would have to conclude this isn’t enough losing to justify breaking the norm. But if not 10%, how do I know what number provides the justification? And do I actually believe that after reaching some number, it would switch from barred to permitted?)
How does this rule fare in the two cases I’m really interested in?
The norm of “don’t support an AI company” seems pretty thoroughly broken, now that there are many AI companies with tens of thousands of collective employees and hundreds of billions of dollars in investment. It also doesn’t seem like there’s some other group holding back on supporting the least-responsible AI companies who we’ll be defecting against if we support the most responsible AI companies. So I’m not sure what’s left of this one under the “don’t break unbroken norms” hypothesis - although I notice this would also justify taking the slightly-less-brutal concentration camp guard job.
(in order to turn down the concentration camp guard job, I think you have to strongly commit to acting as if everyone would follow your rule, even though you know it’s not true - but then how do we handle the military disarmament example?)
The political action campaign is slightly tougher. On the one hand, “don’t use mass politics to achieve your goals” is an impressively non-functioning norm! It’s hard to think of less well-functioning norms than this! On the other, if the Pause AI movement ended up exactly as scummy as the scummiest existing political movements (which tbc I don’t think they’re anywhere close to), then by definition they’re not breaking a norm - we posited some existing movements are already that scummy! - but it still seems bad. Maybe the relevant threshold is “don’t do worse than average”? Or “don’t do worse than your exact political enemies, who in this case are Leading The Future, Chris Lehane, and Marc Andreessen, so you’re probably fine?”
I think this is all defensible, and right now I don’t think either group is violating its particular deontological bar.
Of course, the Early Christian Strategy is to ignore all of this and do the right thing in every case, including unilateral military disarmament. But they can’t keep getting away with it, can they?