Will AI treat us as its “parents”?

It seems quite unlikely.

One hope we’ve heard about AI is that it might treat humanity well because it views us as its “parents.” Unfortunately, this hope seems misplaced.

For one thing, filial love and responsibility seem highly contingent on the details of our evolutionary history.

Almost all mammals and birds care for their young, but in only a few species — including humans — do children care for their parents. Filial responsibility isn’t even universal among primates, let alone in the animal kingdom more broadly. AIs created with gradient descent may have even less in common with humans than that, since AIs don’t have any evolutionary or anatomical connection to humans.

In the case of humans, filial responsibility is strongly correlated with cooperative breeding systems, in which grown offspring stay with their families and help care for their siblings and other extended family members.

There’s a lot that went into human beings developing care for their parents:

Being mammals, hominids invest a lot in their children.
Because of the size and cost of our brains, hominids have a much longer childhood than most other mammals, and thus invest even more in their children.
Hominids benefit from large group structures, for a variety of reasons:
- Defense against large predators
- Coordinated hunting of large prey, and sharing of other perishable food
- The opportunity to learn tool use and other skills by imitation
Before reaching their prime, hominids have a significant ability to help others, such as by providing childcare or doing other forms of basic work.
Old hominids also have the ability to care for children, especially by passing on vital knowledge.
Thus, hominids who cared for their parents had a genetic edge, either by indirectly helping their siblings or by having grandparents who could, in turn, help their grandchildren.
Cultures that promoted filial responsibility also had an edge, for the same reason.

None of these things is likely to be true of AI.

And even if all of them were true, it might not be enough in practice, since any number of other factors could turn out to matter too, such as chaotic variation in the ways AIs reflect on themselves. And, again, filial responsibility is overwhelmingly not the default in the animal kingdom.

One way people imagine AI might acquire a sense of filial responsibility is that it’s trained on a giant corpus of human data, and it interacts with humans a great deal, so perhaps human preferences will “rub off on it” somehow?

We don’t expect this to work. We expect the AI’s preferences to be related to human preferences somehow, but in a tangential, strange, and complicated way — as in the discussion at the end of Chapter 4, where we walk through worlds with more (and increasingly realistic) amounts of complication, in the link between human preferences and AI preferences.

It would probably be bad if they did.

If, against all odds, something like filial responsibility grew inside an AI for one reason or another, we would probably be in a lot of trouble.

An AI can be smart enough to understand exactly what humans mean by “filial responsibility,” while having its own very different version of filial responsibility that it cares about.

Humans were “trained” by natural selection to maximize our reproductive fitness. But nearly all of the things we care about are correlates of fitness — we care little to none about fitness itself.

Similarly, an AI encouraged to “love its parents” would, at best,probably end up with complicated correlates of filial responsibility.

An AI could care deeply about its creators…but not in a way that prizes our subjective experience. In the language of Chapter 4, even “one simple complication” results in versions of “caring about us” that look like freezing us in amber, or keeping humans alive against their will, or preventing us from reproducing and giving the final generation of humans a modestly comfortable environment while the AI takes the rest of the universe for itself. Or something a whole lot weirder than that.

It doesn’t seem possible to predict what the actual outcome would be. But we would expect it to be — if anything — even stranger and less appealing than these options.^[1]

Again, see “Won’t AIs care at least a little about humans?” for related discussion. ↩︎