Humans don't follow anything like a utility function, which is a first problem, so you're asking the AI to construct something that isn't there. Then you have to knit this together into a humanity utility function, which is very non trivial (this is one feeble and problematic way of doing this: http://lesswrong.com/r/discussion/lw/8qb/cevinspired_models/).
The other problem is that you haven't actually solved many of the hard problems. Suppose the AI decides to kill everyone, then replay, in an endless loop, the one upload it has, having a marvellous experience. Why would it not do that? We want the AI to correctly balance our higher order preferences (not being reduced to a single mindless experience) with our lower order preferences (being happy). But that desire is itself a higher order preference - it won't happen unless the AI already decides that higher order preferences trump lower ones.
And that was one example I just thought of. It's not hard to come up with "the AI does something stupid in this model (eg: replaces everyone with chatterbots that describe their ever increasing happiness and fulfilment) that is compatible with the original model but clearly stupid - clearly stupid to our own judgement, though, not to the AIs.
You may object that these problems won't happen - but you can't be confident of this, as you haven't defined your solution formally, and are relying on common sense to reject those pathological solutions. But nowhere have you assumed the AI has common sense, or how it will use it. The more details you put in your model, I think, the more the problems will become apparent.
Thank you for the thoughtful reply!
Deducing the correct utility of a utility maximiser is one thing (which has a low level of uncertainty, higher if the agent is hiding stuff).
In the white-box approach it can't really hide. But I guess it's rather tangential to the discussion.
Assigning a utility to an agent that doesn't have one is quite another... Humans don't follow anything like a utility function, which is a first problem, so you're asking the AI to construct something that isn't there.
What do you mean by "follow a utility function"?...
To construct a friendly AI, you need to be able to make vague concepts crystal clear, cutting reality at the joints when those joints are obscure and fractal - and them implement a system that implements that cut.
There are lots of suggestions on how to do this, and a lot of work in the area. But having been over the same turf again and again, it's possible we've got a bit stuck in a rut. So to generate new suggestions, I'm proposing that we look at a vaguely analogous but distinctly different question: how would you ban porn?
Suppose you're put in change of some government and/or legal system, and you need to ban pornography, and see that the ban is implemented. Pornography is the problem, not eroticism. So a lonely lower-class guy wanking off to "Fuck Slaves of the Caribbean XIV" in a Pussycat Theatre is completely off. But a middle-class couple experiencing a delicious frisson when they see a nude version of "Pirates of Penzance" at the Met is perfectly fine - commendable, even.
The distinction between the two case is certainly not easy to spell out, and many are reduced to saying the equivalent of "I know it when I see it" when defining pornography. In terms of AI, this is equivalent with "value loading": refining the AI's values through interactions with human decision makers, who answer questions about edge cases and examples and serve as "learned judges" for the AI's concepts. But suppose that approach was not available to you - what methods would you implement to distinguish between pornography and eroticism, and ban one but not the other? Sufficiently clear that a scriptwriter would know exactly what they need to cut or add to a movie in order to move it from one category to the other? What if the nude "Pirates of of Penzance" was at a Pussycat Theatre and "Fuck Slaves of the Caribbean XIV" was at the Met?
To get maximal creativity, it's best to ignore the ultimate aim of the exercise (to find inspirations for methods that could be adapted to AI) and just focus on the problem itself. Is it even possible to get a reasonable solution to this question - a question much simpler than designing a FAI?