A putative new idea for AI control; index here.

For anyone but an extreme total utilitarian, there is a great difference between AIs that would eliminate everyone as a side effect of focusing on their own goals (indifferent AIs) and AIs that would effectively eliminate everyone through a bad instantiation of human-friendly values (false-friendly AIs). Examples of indifferent AIs are things like paperclip maximisers, examples of false-friendly AIs are "keep humans safe" AIs who entomb everyone in bunkers, lobotomised and on medical drips.

The difference is apparent when you consider multiple AIs and negotiations between them. Imagine you have a large class of AIs, and that they are all indifferent (IAIs), except for one (which you can't identify) which is friendly (FAI). And you now let them negotiate a compromise between themselves. Then, for many possible compromises, we will end up with most of the universe getting optimised for whatever goals the AIs set themselves, while a small portion (maybe just a single galaxy's resources) would get dedicated to making human lives incredibly happy and meaningful.

But if there is a false-friendly AI (FFAI) in the mix, things can go very wrong. That is because those happy and meaningful lives are a net negative to the FFAI. These humans are running dangers - possibly physical, possibly psychological - that lobotomisation and bunkers (or their digital equivalents) could protect against. Unlike the IAIs, which would only complain about the loss of resources to the FAI, the FFAI finds the FAI's actions positively harmful (and possibly vice versa), making compromises much harder to reach.

And the compromises reached might be bad ones. For instance, what if the FAI and FFAI agree on "half-lobotomised humans" or something like that? You might ask why the FAI would agree to that, but there's a great difference to an AI that would be friendly on its own, and one that would choose only friendly compromises with a powerful other AI with human-relevant preferences.

Some designs of FFAIs might not lead to these bad outcomes - just like IAIs, they might be content to rule over a galaxy of lobotomised humans, while the FAI has its own galaxy off on its own, where its humans take all these dangers. But generally, FFAIs would not come about by someone designing a FFAI, let alone someone designing a FFAI that can safely trade with a FAI. Instead, they would be designing a FAI, and failing. And the closer that design got to being FAI, the more dangerous the failure could potentially be.

So, when designing an FAI, make sure to get it right. And, though you absolutely positively need to get it absolutely right, make sure that if you do fail, the failure results in a FFAI that can safely be compromised with, if someone else gets out a true FAI in time.

New Comment
12 comments, sorted by Click to highlight new comments since: Today at 9:05 AM

Alternatively, suppose that there is a parliament of all IAIs vs. a parliament of all FFAIs. It could be that the false-friendly AIs, who are each protecting some aspect of humanity, end up doing an ok job. Not optimal, but the AI who wants humans to be safe and the AI that wants humans to have fun, and the AI that wants humans to spread across the galaxy all together could lead to something not awful. The IAI parliament on the other hand just leads to the humans getting turned into resources for whichever of them can most convenient make use of our matter.

Notice that adding IAIs to the FFAIs does nothing more (according to many ways of resolving disagreements) than reducing the share of resources humanity gets.

But counting on a parliament of FFAIs to be finely balanced to get FAI out of it, without solving FAI along the way... seems a tad optimistic. You're thinking of "this FFAI values human safety, this one values human freedom, they will compromise on safety AND freedom". I'm thinking they will compromise on some lobotomy-bunker version of safety while running some tiny part of the brains to make certain repeated choices that technically count as "freedom" according to the freedom-FFAI's utility.

I'm just brainstorming in the same vein as these posts, of course, so consider the epistemic status of these comments to be extremely uncertain. But, in the limit, if you have a large number of AIs (thousands, or millions, or billions) who each optimize for some aspect that humans care about, maybe the outcome wouldn't be terrible, although perhaps not as good as one truly friendly AI. The continuity of experience AI could compromise with the safety AI and freedom AI and "I'm a whole brain experiencing things" AI and the "no tricksies" AI to make something not terrible.

Of course, people don't care about so many aspects with equal weights, so if they all got equal weight, maybe the most likely failure mode is that something people only care about a tiny amount (e.g. not stepping on cracks in the sidewalk) gets equal weight with something people care about a lot (e.g. experiencing genuine love for another human) and everything gets pretty crappy. On the other hand, maybe there are many things that can be simultaneously satisfied, so you end up living in a world with no sidewalk-cracks and where you are immediately matched with plausible loves of your life, and while it may not be optimal, it may still be better than what we've got going on now.

I'll think about it. I don't think it will work, but there might be an insight there we can use.

but there's a great difference to an AI that would be friendly on its own, and one that would choose only friendly compromises with a powerful other AI with human-relevant preferences.

An AI must compromise with the universe and only implement something physically possible. If it's going to make sure this compromise is friendly, why wouldn't it make a friendly compromise with an FFAI?

why wouldn't it make a friendly compromise with an FFAI?

Because that's an extra constraint: universe AND FFAI. The class of AIs that would be FAI with the universe, is larger than the class that would be FAI with the universe and an FFAI to deal with.

To pick a somewhat crude example, imagine an AI that maximises the soft-minimum of two quantities: human happiness and human preferences. It turns out each quantity is roughly equivalent in difficulty of satisfying (ie not too many order of magnitudes between them), so this is a FAI in our universe.

However, add a FFAI that hates human preferences and loves human happiness. Then the compromise might be on a very high happiness, which the previous FAI can live with (it was only a soft minimum, not a hard minimum).

Or maybe this is a better way of formulating things: there are FAIs, and AIs which act as FAIs given the expected conditions of the universe. It's the second category that might be very problematic in negotiations.