"Would you kill babies if it was the right thing to do? If no, under what circumstances would you not do the right thing to do? If yes, how right would it have to be, for how many babies?"
—horrible job interview question

Swapping hats for a moment, I'm professionally intrigued by the decision theory of "things you shouldn't do even if they seem to be the right thing to do".

Suppose we have a reflective AI, self-modifying and self-improving, at an intermediate stage in the development process. In particular, the AI's goal system isn't finished—the shape of its motivations is still being loaded, learned, tested, or tweaked.

Yea, I have seen many ways to screw up an AI goal system design, resulting in a decision system that decides, given its goals, that the universe ought to be tiled with tiny molecular smiley-faces, or some such. Generally, these deadly suggestions also have the property that the AI will not desire its programmers to fix it. If the AI is sufficiently advanced—which it may be even at an intermediate stage—then the AI may also realize that deceiving the programmers, hiding the changes in its thoughts, will help transform the universe into smiley-faces.

Now, from our perspective as programmers, if we condition on the fact that the AI has decided to hide its thoughts from the programmers, or otherwise act willfully to deceive us, then it would seem likely that some kind of unintended consequence has occurred in the goal system. We would consider it probable that the AI is not functioning as intended, but rather likely that we have messed up the AI's utility function somehow. So that the AI wants to turn the universe into tiny reward-system counters, or some such, and now has a motive to hide from us.

Well, suppose we're not going to implement some object-level Great Idea as the AI's utility function. Instead we're going to do something advanced and recursive—build a goal system which knows (and cares) about the programmers outside. A goal system that, via some nontrivial internal structure, "knows it's being programmed" and "knows it's incomplete". Then you might be able to have and keep the rule:

"If [I decide that] fooling my programmers is the right thing to do, execute a controlled shutdown [instead of doing the right thin

...

Carrie_Toombs

Carrie_Toombs

Carrie_Toombs

Carrie_Toombs