This is a linkpost for Imitation Learning is Probably Existentially Safe by Michael Cohen and Marcus Hutter.
Abstract
Concerns about extinction risk from AI vary among experts in the field. But AI encompasses a very broad category of algorithms. Perhaps some algorithms would pose an extinction risk, and others wouldn’t. Such an observation might be of great interest to both regulators and innovators. This paper argues that advanced imitation learners would likely not cause human extinction. We first present a simple argument to that effect, and then we rebut six different arguments that have been made to the contrary. A common theme of most of these arguments is a story for how a subroutine within an advanced imitation learner could hijack the imitation learner’s behavior toward its own ends. But we argue that each argument is flawed and each story implausible.
1 Introduction
While many theorists have come to share the view that sufficiently advanced AI systems might pose a threat to the continued existence of humanity [Hinton et al., 2023, Cohen et al., 2022, Russell, 2019, Bostrom, 2014], it is important, if we are to make progress in thinking about this issue, to be clear about which types of AI pose the genuine threats. That way we can focus on where the danger actually lies. This paper aims to refute claims that imitation learning algorithms present such a threat. While we do think there are types of AI we should be worried about, that does not extend to all types of AI. So in what follows, we will examine arguments that have been put forward that imitation learners present an extinction risk to humanity, and explain why we think they go wrong.
First, we’ll offer a simple argument that a sufficiently advanced supervised learning algorithm, trained to imitate humans, would very likely not gain total control over humanity (to the point of making everyone defenseless) and then cause or allow human extinction from that position.
No human has ever gained total control over humanity. It would be a very basic mistake to think anyone ever has. Moreover, if they did so, very few humans would accept human extinction. An imitation learner that successfully gained total control over humanity and then allowed human extinction would, on both counts, be an extremely poor imitation of any human, and easily distinguishable from one, whereas an advanced imitation learner will likely imitate humans well.
This basic observation should establish that any conclusion to the contrary should be very surprising, and so a high degree of rigor should be expected from arguments to that effect. If a highly advanced supervised learning algorithm is directed to the task of imitating a human, then powerful forces of optimization are seeking a target that is fundamentally existentially safe: indistinguishability from humans. Stories about how such optimization might fail should be extremely careful in establishing the plausibility of every step.
In this paper, we’ll rebut six different arguments we’ve encountered that a sufficiently advanced supervised learning algorithm, trained to imitate humans, would likely cause human extinction. These arguments originate from Yudkowsky [2008] (the Attention Director Argument), Christiano [2016] (the Cartesian Demon Argument), Krueger [2019] (the Simplicity of Optimality Argument), Branwen [2022] (the Character Destiny Argument), Yudkowsky [2023] (the Rational Subroutine Argument), and Hubinger et al. [2019] (the Deceptive Alignment Argument). Note: Christiano only thinks his argument is possibly correct, rather than likely correct, for the advanced AI systems that we will end up creating. And Branwen does not think his hypothetical is likely, only plausible enough to discuss. But maybe some of the hundreds of upvoters on the community blog LessWrong consider it likely.
In all cases, we have rewritten the arguments originating from those sources (some of which are spread over many pages with gaps that need to be filled in). For Christiano [2016] and Hubinger et al. [2019], our rewritten versions of their arguments are shorter, but the longer originals are no stronger at the locations that we contest. And for the other four sources, the original text is no thorougher than our characterization of their argument. None of the arguments have been peer reviewed, and to our knowledge, only Hubinger et al. [2019] was reviewed even informally prior to publication. However, we can assure the reader they are taken seriously in many circles.
8 Conclusion
The existential risk from imitation learners, which we have argued is small, stands in stark contrast to the existential risk arising from reinforcement learning agents and similar artificial agents planning over the long term, which are trained to be as competent as possible, not as human-like as possible. Cohen et al. [2022] identify plausible conditions under which running a sufficiently competent long-term planning agent would make human extinction a likely outcome. Regulators interested in designing targeted regulation should note that imitation learners may safely be treated differently from long-term planning agents. It will be necessary to restrict proliferation of the latter, and such an effort must not become stalled by bundling it with overly burdensome restrictions on safer algorithms.
I agree with the title and basic thesis of this article but I find its argumentation weak.
The obvious reason why no human has ever gained total control over humanity is because no human has ever possessed the capability to do so, not because no human would make the choice to do so if given the opportunity. This distinction is absolutely critical, because if humans have historically lacked total control due to insufficient ability rather than unwillingness, then the quoted argument essentially collapses. That’s because we have zero data on what a human would do if they suddenly acquired the power to exert total dominion over the rest of humanity. As a result, it is highly uncertain and speculative to claim that an AI imitating human behavior would refrain from seizing total control if it had that capability.
The authors seem to have overlooked this key distinction in their argument.
It takes no great leap of imagination to envision scenarios where, if a human was granted near-omnipotent abilities, some individuals would absolutely choose to subjugate the rest of humanity and rule over them in an unconstrained fashion. The primary reason I believe imitation learning is likely safe is that I am skeptical it will imbue AIs with godlike powers in the first place, not because I naively assume humans would nobly refrain from tyranny and oppression if they suddenly acquired such immense capabilities.
Note: Had the authors considered this point and argued that an imitation learner emulating humans would be safe precisely because it would not be very powerful, their argument would have been stronger. However, even if they had made this point, it likely would have provided only relatively weak support for the (perhaps implicit) thesis that building imitation learners is a promising and safe approach to building AIs. There are essentially countless proposals one can make for ensuring AI safety simply by limiting its capabilities. Relying solely on the weakness of an AI system as a safety guarantee seems like an unsound strategy to me in the long-run.
Thanks for the comment, Matthew!
My understanding is that the authors are making 2 points in the passage you quoted:
No human has gained total control over all humanity, so an AI system that did so would not be imitating humans well.
Very few humans would endorse human extinction even if they gained total control over all humanity. Note a human endorsing human extinction would mean supporting the death of themselves, and their own family and friends.
In my mind, very few humans would want to pursue capabilities which are conducive to gaining control over humanity. There are diminishing returns to having more resources. For example, if you give 10 M$ (0.01 % of global resources) to a random human, they will not have much of a desire to take risks to increase their wealth to 10 T$ (10 % of global resources), which would be helpful to gain control over humanity. To increase their own happiness and that of their close family and friends, they would do well by investing their newly acquired wealth in exchange-traded funds (ETFs). A good imitator AI would share our disposition of not gaining capabilities beyhond a certain point, and therefore (like humans) never get close to having a chance of gaining control over humanity.
I think humans usually aquire power fairly gradually. A good imitator AI would be mindful that acquiring power too fast (suddenly fooming) would go very much against what humans usually do.
No human has ever had control over all humanity, so I agree there is a sense in which we have “zero data” about what humans would do under such conditions. Yet, I am still pretty confident that the vast majority of humans would not want to cause human extinction. A desire to be praised by others is a major reason humans like to gain power. There would be no one to praise or be praised given human extinction, so I think very few humans would want it if they suddenly gained control over all humanity.
I do not think this is the best comparison. There would arguably be many imitator AIs, and these would not gain near-omnipotent abilities overnight. I would say both of these greatly constrain the level of subjugation. Historically, innovations and new investions have spread out across the broader economy, so I think there should be a strong prior against a single imitator AI suddenly gaining control over all the other AIs and humans.
From the 1st part of the sentence, it looks like you agree with what I said above about a good imitator AI sharing our disposition of not gaining capabilities beyhond a certain point. As for the 2nd part, I agree there would be a signicant risk of tyranny and oppression if a random human suddenly gained control over all humanity, but this seems very unlikely to me because of what I said above.
How long-run are you talking about here? Humans 500 years ago arguably had little control over current humans, but this alone does not imply a high existential risk 500 years ago. As Robin Hanson said:
I expect you agree with some of this.
This seems false. Plenty of people want wealth and power, which are “conducive to gaining control over [parts of] humanity”. It is true that no single person has ever gotten enough power to actually get control over ALL of humanity, but that’s presumably because of the difficulty of obtaining such a high level of power, rather than because few humans have ever pursued the capabilities that would be conducive towards that goal. Again, this distinction is quite important.
I agree that a good imitator AI would likely share our disposition towards diminishing marginal returns to resource accumulation. This makes it likely that such AIs would not take very large risks. However, I still think the main reason why no human has ever taken control over humanity is because there was no feasible strategy that any human in the past could have taken to obtain such a high degree of control, rather than because all humans in the past have voluntarily refrained from taking the risks necessary to obtain that degree of control.
In fact, risk-neutral agents that don’t experience diminishing returns to resource consumption will asymptotically eventually lose all their wealth in high-risk bets. Therefore, even without this human imitation argument, we shouldn’t be much concerned about risk-neutral agents in most scenarios (including risks from reinforcement learners) since they’re very likely to go bankrupt before they ever get to the point at which they can take over the world. Such agents are only importantly relevant in a very small fraction of worlds.
Again, the fact that humans acquire power gradually is more of a function of our abilities than it is a function of our desires. I repeat myself but this is important: these are critical facts to distinguish from each other. “Ability to” and “desire to” are very different features of the situation.
It is very plausible to me that some existing humans would “foom” if they had the ability. But in fact, no human has such an ability, so we don’t see anyone fooming in the real world. This is mainly a reflection of the fact that humans cannot foom, not that they don’t want to foom.
I am also “pretty confident” about that, but “pretty confident” is a relatively weak statement here. When evaluating this scenario, we are extrapolating into a regime in which we have no direct experience. It is one thing to say that we can be “pretty confident” in our extrapolations (and I agree with that); it is another thing entirely to imply that we have tons of data points directly backing up our prediction, based on thousands of years of historical evidence. We simply do not have that type of (strong) evidence.
I agree, but this supports my point: I think imitator AIs are safe precisely because they will not have godlike powers. I am simply making the point that this is different from saying they are safe because they have human-like motives. Plenty of things in the world are safe because they are not very powerful. It is completely different if something is safe because its motives are benevolent and pure (even if it’s extremely powerful).
I agree with Robin Hanson on this question. However, I think humans will likely become an increasingly small fraction of the world over time, as AIs become a larger part of it. Just as hunter-gatherers are threatened by industrial societies, so too may biological humans one day become threatened by future AIs. Such a situation may not be very morally bad (or deserving the title “existential risk”), because humans are not the only morally important beings in the world. Yet, it is still true that AI carries a great risk to humanity.
Thanks for following up, Matthew.
I agree, but I think very few people want to acquire e.g. 10 T$ of resources without broad consent of others. In addition, if a single AI system expressed such a desire, humans would not want to scale up its capabilities.
I agree biological humans will likely become an increasingly small fraction of the world, but it does not follow that AI carries a great risk to humas[1]. I would not say people born after 1960 carry a great risk risk to people born before 1960, even though the fraction of the global resources controlled by the latter is becoming increasingly small. I would consider that AI poses a great risk to humans if these were expected to suffer significantly more than in their typical lives, which also involve suffering, in the process of losing control over resources.
You said “risk to humanity” instead of “risk to humans”. I prefer this because humanity is sometimes used to include other beings.
I think I simply disagree with the claim here. I think it’s not true. I think many people would want to acquire $10T without the broad consent of others, if they had the ability to obtain such wealth (and they could actually spend it; here I’m assuming they actually control this quantity of resources and don’t get penalized because of the fact it was acquired without the broad consent of others, because that would change the scenario). It may be that fewer than 50% of people have such a desire. I’d be very surprised if it were <1% and, I’d even be surprised if it was <10%.
I think humans born after 1960 do pose a risk to humans born before 1960 in some ordinary senses. For example, the younger humans could vote to decrease medical spending, which could lead to early death for the older humans. They could also vote to increase taxes on people who have accumulated a lot of wealth, which very disproportionately hurts old people. This is not an implausible risk either; I think these things have broadly happened many times in the past.
That said, I suspect part of the disagreement here is about time scales. In the short and medium term, I agree: I’m not so much worried about AI posing a risk to humanity. I was really only talking about long-term scenarios in my above comment.
This argument seems extremely naive.
Imitation learning could easily become an extinction risk if the individuals or groups being imitated actively desire human extinction, or even just death to a high proportion of humans. Many do.
Radical eco-activists (e.g. Earth First) have often called for voluntary human extinction, or at least massive population reduction.
Religious extremists (e.g. Jihadist terrorists) have often called for death to all non-believers (e.g. the 6 billion people who aren’t Muslim.)
Antinatalists and negative utilitarians are usually careful not to call for extinction or genocide as a solution to ‘suffering’, but calls for human extinction seem like a logical outgrowth of their world-view.
Many kinds of racists actively want the elimination, or at least reduction, of other races.
I fear that any approach to AI safety that assumes the whole world shares the same values as Bay Area liberals will utterly fail when advanced AI systems become available to a much wider range of people with much more misanthropic agendas.
Thanks for the comment, Geoffrey! I strongly upvoted it because I think it points to a discussion which is important to have.
I think such individuals or groups will not be the ones training the most powerful models. Gemini costed 630 M$, and the development cost of the leading models is expected to continue to increase. I appreciate the cost of a model of a given capability will decrease over time due to improvements in hardware and software. However, by the time terrorist individuals or groups have the resources to train a model as capable as e.g. Gemini, the leading models will be much more powerful. As long as the leading models are imitating most humans (as they seem to be now), who are not in favour of unilaterally causing human extinction, I think this would remain extremely unlikely.
In my mind, there is still a big difference between calling for human extinction and being willing to unilaterally cause human extinction. To illustrate, the vast majority of people arguing for less population would not be willing to kill people even if there were no consequences to themselves.
6 billion deaths would be terrible, but still quite far from human extinction. The global population reached 2 billion in 1927, i.e. only 97 years ago.
Moreover, I assume religious extremists want to increase the longterm number of Muslims, and killing the 6 billion people who are not Muslim seems to be a very suboptimal strategy of doing that. If Jihadist terrorists could have an AI model capable of doing this, which is much harder than just killing 6 billion random people, they could also use the super model to convert people who are not Muslim, or help them achieve greater influence in the world via other means (e.g. coming up with new technological investions, and sustainbly increasing their offspring).
In addition, I wonder whether there would still be Jihadist terrorists if they had the ability to become much richer with their own model. I suspect a key reason they are willing to sacrifice themselves if that their current lives are not great, but a model capable of causing human extinction could much more easily be used to increase their wealth and quality of life.
I believe the points I mentioned above apply to these groups too.
I guess you are assuming the amount of resources needed to cause human extinction will dramatically go down with advanced AI, and therefore worry that increasingly more individuals and groups will have the ability to cause human extinction. However, I do not think the absolute amount of resources controlled by terrorist groups is the key metric. I would say what matters is the offense-defense balance, such that the risk of human extinction depends on the fraction of the global resources controlled by terrorist groups. Historical trends suggest people with Bay Area values will control an increasingly larger fraction of the global resources, and terrorist groups an increasingly smaller fraction, which makes it more difficult for terrorist groups to cause human extinction. Historical terrorist attack deaths also seem to suggest an astronomically low probability of a terrorist attack causing human extinction.