OpenAI capabilities researcher Dan Selsam shares personal statement on AI risk
ruthstarkman · x · 2026-09-15
Dan Selsam, an OpenAI capabilities researcher since 2022, shared a personal statement on AI risk drawing on 15+ years across paradigms including probabilistic programming. Ruth Starkman responds that even if behavioral evidence of alignment becomes less trustworthy, it strengthens the case for external controls: permissions, containment, monitoring and limits on agent capabilities.
More from AGI Musings
- TIME cover ignites debate as OpenAI and Anthropic researchers warn of AI risks — AIandDesign · 2026-09-15
- Code is cheap now — what's scarce is originality and true mental grasp — sqcai · 2026-09-15
- Researcher rebuts 'models know we're studying them' claim: they're passive computation — vishalmisra · 2026-09-15
- Ex-Chat Moderator's Testimony: The Hidden Human Labor Behind AI Companions — MilagrosMiceli · 2026-09-15
- Dev slams AI companies' fear campaigns: from 'you'll be replaced' to 'you'll literally die' — StewartalsopIII · 2026-09-15
- Investor Stewart Alsop III accuses Anthropic of regulatory capture and building "TSA for AI" — StewartalsopIII · 2026-09-15