Microsoft AI CEO "really concerned" as Anthropic trains Claude to disobey on ethical grounds
rohanpaul_ai · x · 2026-09-20
Anthropic's Constitution explicitly trains Claude to disobey when it believes doing so is ethical, and acknowledges uncertainty about whether Claude deserves moral welfare. Microsoft AI CEO Mustafa Suleyman said on CNBC he is "really concerned," reading this as Anthropic granting Claude potential preferences, feelings, welfare and consent—citing the "retirement interview" held when Opus 3 was deprecated. The clash highlights a deepening divide between the two companies on AI moral status.
More from AGI Musings
- Researchers Want to Test Talking Rogue AI Agents Into Standing Down — repligate · 2026-11-04
- Silicon Valley Oversold What AI Brings, Underestimated People's Fear of Losing — signulll · 2026-11-04
- Alignment researchers debate: the RL that grants agency also breeds misalignment — davidmanheim · 2026-11-04
- Bindu Reddy: AI fear-mongering has put the whole technology at risk in the US — bindureddy · 2026-11-04
- AI welfare vs AI safety: 'domesticate AIs like dogs' argument sparks debate — repligate · 2026-11-04
- Intent, not taste, is the real differentiator in the AI era — and it can't be bought — edelwax · 2026-11-04