In this essay Paul Christiano proposes a definition of "AI alignment" which is more narrow than other definitions that are often employed. Specifically, Paul suggests defining alignment in terms of the motivation of the agent (which should be, helping the user), rather than what the agent actually does. That is, as long as the agent "means well", it is aligned, even if errors in its assumptions about the user's preferences or about the world at large lead it to actions that are bad for the user.
Rohin Shah's comment on the essay (which I believe is endorsed by Paul) reframes it as a particular way to decompose the AI safety problem. An often used decomposition is "definition-optimization": first we define what it means for an AI to be safe, then we understand how to implement a safe AI. In contrast, Paul's definition of alignment decomposes the AI safety problem as "motivation-competence": first we learn how to design AIs with good motivations, then we learn how to make them competent. Both Paul and Rohin argue that the "motivation" is the urgent part of the problem, the part on which technical AI safety research should focus.
In contrast, I will argue that the "motivation-competence" decomposition is not as useful as Paul and Rohin believe, and the "definition-optimization" decomposition is more useful.
The thesis behind the "motivation-competence" decomposition implicitly assumes a linear, one-dimensional scale of competence. Agents with good motivations and subhuman competence might make silly mistakes but are not catastrophically dangerous (since they are subhuman). Agents with good motivations and superhuman competence will only do mistakes that are "forgivable" in the sense that, our own mistakes would be as bad or worse. Ergo (the thesis concludes), good motivations are sufficient to solve AI safety.
However, in reality competence is multi-dimensional. AI systems can have subhuman skills in some domains and superhuman skills in other domains, as AI history showed time and time again. This opens the possibility of agents that with "well intentioned" mistakes that take the form of sophisticated plans that are catastrophic for the user. Moreover, there might be limits to the agent's knowledge about certain questions (such as, the user's preferences) that are inherent in the agent's epistemology (more on this below). Given such limits, the agent's competence becomes systematically lopsided. Furthermore, the elimination of such limits is as a large part of the "definition" part in the "definition-optimization" framing that the thesis rejects.
As a consequence of the multi-dimensional natural of competence, the difference between "well intentioned mistake" and "malicious sabotage" is much less clear than naively assumed, and I'm not convinced there is a natural way to remove the ambiguity. For example, consider a superhuman AI Alpha subject to an acausal attack. In this scenario, some agent Beta in the "multiverse" (= prior) convinces Alpha that Alpha exists in a simulation controlled by Beta. The simulation is set up to look like the real Earth for a while, making it a plausible hypothesis. Then, a "treacherous turn" moment arrives in which the simulation diverges from Earth, in a way calculated to make Alpha take irreversible actions that are beneficial for Beta and disastrous for the user.
In the above scenario, is Alpha "motivation-aligned"? We could argue it is not, because it is running the malicious agent Beta. But we could also argue it is motivtion-aligned, it just makes the innocent mistake of falling for Beta's trick. Perhaps it is possible to clarify the concept of "motivation" such that in this case, Alpha's motivations are considered bad. But, such a concept would depend in complicated ways on the agent's internals. I think that this is a difficult and unnatural approach, compared to "definition-optimization" where the focus is not on the internals but on what the agent actually does (more on this later).
The possibility of acausal attacks is a symptom of the fact that, environments with irreversible transitions are usually not learnable (this is the problem of traps in reinforcement learning, that I discussed for example here and here), i.e. it is impossible to guarantee convergence to optimal expected utility without further assumptions. When we add preference learning to the mix, the problem gets worse because now even if there are no irreversible transitions, it is not clear the agent will converge to optimal utility. Indeed, depending on the value learning protocol, there might be uncertainties about the user's preferences that the agent can never resolve (this is an example of what I meant by "inherent limits" before). For example, this happens in CIRL (even if the user is perfectly rational, this happens because the user and the AI have different action sets).
These difficulties with the "motivation-competence" framing are much more natural to handle in the "definition-optimization" framing. Moreover, the latter already produced viable directions for mathematical formalization, and the former has not (AFAIK). Specifically, the mathematical criteria of alignment I proposed are the "dynamic subjective regret bound" and the "dangerousness bound". The former is a criterion which simultaneous guarantees motivation-alignment and competence (as evidence that this criterion can be satisfied, I have the Dialogic Reinforcement Learning proposal). The latter is a criterion that doesn't guarantee competence in general, but guarantees specifically avoiding catastrophic mistakes. This makes it closer to motivation-alignment compated to subjective regret, but different in important ways: it refers to the actual things that agent does, and the ways in which these things might have catastrophic consequences.
In summary, I am skeptical that "motivation" and "competence" can be cleanly separately in a way that is useful for AI safety, whereas "definition" and "optimization" can be so separated: for example the dynamic subjective regret bound is a "definition" whereas dialogic RL and putative more concrete implementations thereof are "optimizations". My specific proposals might have fatal flaws that weren't discovered yet, but I believe that the general principle of "definition-optimization" is sound, while "motivation-competence" is not.
I decided that the answer deserves its own post.
When I say an AI A is aligned with an operator H, I mean:
The “alignment problem” is the problem of building powerful AI systems that are aligned with their operators.
This is significantly narrower than some other definitions of the alignment problem, so it seems important to clarify what I mean.
In particular, this is the problem of getting your AI to try to do the right thing, not the problem of figuring out which thing is right. An aligned AI would try to figure out which thing is right, and like a human it may or may not succeed.
Analogy
Consider a human assistant who is trying their hardest to do what H wants.
I’d say this assistant is aligned with H. If we build an AI that has an analogous relationship to H, then I’d say we’ve solved the alignment problem.
“Aligned” doesn’t mean “perfect:”
I use alignment as a statement about the motives of the assistant, not about their knowledge or ability. Improving their knowledge or ability will make them a better assistant — for example, an assistant who knows everything there is to know about H is less likely to be mistaken about what H wants — but it won’t make them more aligned.
(For very low capabilities it becomes hard to talk about alignment. For example, if the assistant can’t recognize or communicate with H, it may not be meaningful to ask whether they are aligned with H.)
Clarifications
Postscript on terminological history
I originally described this problem as part of “the AI control problem,” following Nick Bostrom’s usage in Superintelligence, and used “the alignment problem” to mean “understanding how to build AI systems that share human preferences/values” (which would include efforts to clarify human preferences/values).
I adopted the new terminology after some people expressed concern with “the control problem.” There is also a slight difference in meaning: the control problem is about coping with the possibility that an AI would have different preferences from its operator. Alignment is a particular approach to that problem, namely avoiding the preference divergence altogether (so excluding techniques like “put the AI in a really secure box so it can’t cause any trouble”). There currently seems to be a tentative consensus in favor of this approach to the control problem.
I don’t have a strong view about whether “alignment” should refer to this problem or to something different. I do think that some term needs to refer to this problem, to separate it from other problems like “understanding what humans want,” “solving philosophy,” etc.
This post was originally published here on 7th April 2018.
The next post in this sequence will post on Saturday, and will be "An Unaligned Benchmark" by Paul Christiano.
Tomorrow's AI Alignment Sequences post will be the first in a short new sequence of technical exercises from Scott Garrabrant.