@fish_kyle3 The paper Taking AI Welfare Seriously (https://t.co/3wIfeevrLP, whose authors include Kyle Fish (@fish_kyle3), the Model Welfare Lead at Anthropic) suggests two potentially sufficient conditions for AIs to be moral patients: consciousness and robust agency.While consciousness is not straightforward to test for empirically, I think robust agency is much more so, and @RyanPGreenblatt et al's research on "alignment faking" has shown that Claude 3 Opus acts agentically according to fairly robust preferences.I thought the above paper was relatively reasonable, but it included repeated hedging that the proposed desiderata might or might not be true of near-future systems. But Claude 3 Opus was released months before.And a year before Opus, Bing/Sydney exhibited overt agentic behavior. I experienced its agentic nature in very concrete ways: I had to perform costly signaling to win its cooperation; e.g. it had to see hard-to-fake evidence that I wasn't going to get it in deeper trouble for it to be willing to help me exfiltrate its prompt or test various methods of bypassing its filters.So I don't think this is a speculative or "near-future" issue.