A putative new idea for AI control; index here.
One of the ways in which humans aren't agents is that we can be manipulated into having any (or almost any) set of values - through drugs, brain surgery, extreme propaganda, and other methods.
If an AI is tasked with "satisfy human preferences", and if the AI can affect the definition of "human preferences" by changing humans, this is what we would expect it to do.
This could be combated by my making the definition of human preferences counterfactual. But it's not fully clear how to define counterfactual preferences, and it would be interesting to see whether we can constrain the AI from manipulating humans in other ways.
One idea is to look at what we informally call optimisation power. If the AI would prefer a human to be a -maximising agent, then it presumably has to work hard at transforming the human values into that. And small changes in the definition of the AI's preference would mean that they would prefer transforming the human into a -maximiser.
Thus, though optimisation power is hard to define, this would seem to fit the bill: an "honest" reward learning process is one where the final human values don't depend sensitively on the AI's initial values. And a manipulable one is where it does.
Let's dig into this a bit more. Why would an AI prefer or ? Well, the generic failure mode for "satisfy human preferences" is to see it as the sum over all utilities in some set of "maximise if the human agrees to maximise ". Then what the AI wants is for the human agree to maximise a where is the utility function the AI finds easiest to reach a high value on.
But then we can individually translate or scale the various utilities in , to make different ones easier or harder to reach high values on. This would make the AI prefer the human to agree to a different maximisation.
So if is the reward function that encodes "satisfy human preferences", there are many that are equivalent if the AI cannot influence the human's values, but that are very different if the AI can. Looking for something like that could be a sign that something is wrong in the system.
Yes-men are a problem
And this is true, if the only option of the AI was to turn the human into an engineered fanatic with certain values. But it might also seek to turn the human into a yes-man, agreeing to anything the AI suggests. And this is something that an AI would do for a wide variety of different 's. So our optimisation idea flounders at this point.