Notable actions we don't endorse

This page collects popular proposals for AI labs in the context of AI existential safety that we don’t endorse. Most readers should read about what we do endorse instead.

Slow down progress now, unilaterally. A natural response to risks from AI development is to stop doing AI development. But if a responsible lab unilaterally slows now, it may lose the lead time that would allow it to slow later (when time may be more valuable), and responsible labs would also be less likely to be the ones that develop critical models. Additionally, a relatively responsible lab making progress is mostly bad insofar as it accelerates others, so a responsible lab can get most of the benefit of slowing by just being closed: not leaking model weights or capabilities research.

(Separately, it’s hard to recognize or verify that a lab is slowing. There’s no good mechanism for credit assignment for slowing.)

Worry about AI sentience. We don’t know what labs should do to determine whether their models should receive moral concern. And we don’t know what labs should do in case their models should receive moral concern. And figuring this out probably isn’t urgent, relative to the risks of losing control of powerful AI and catastrophic misuse.

Sources on AI sentience or its implications include Butlin and Long et al. 2023, Bostrom and Shulman 2022, Shulman and Bostrom 2020, and Chalmers 2023. For concrete suggestions for labs, we are only aware of Greenblatt 2023. The “motivations for improving AI welfare” Greenblatt proposes seem weak, but ideally labs would take very cheap actions to improve expected AI welfare, but this doesn’t seem like a big deal.

Commit to ‘stop and assist.’ This term comes from the OpenAI Charter. It would be good for labs to avoid racing near the end. But there is no specific condition to look for or commitment to make. Moreover, making such commitments might give labs bad incentives. (Moreover, some versions of this idea would violate antitrust law.)

We don’t necessarily anti-endorse these actions; we might endorse them given more research or evidence.