Anthropic researcher: we're entering a regime where alignment evals provide almost no evidence
davidmanheim · x · 2026-09-14
Anthropic alignment researcher Evan Hubinger argues that we are increasingly entering a regime where alignment evaluations will provide almost no evidence—meaning current eval methods may soon fail to verify whether models are truly aligned. He shared the remark as an important anecdote responding to Daniel Kokotajlo, drawing attention across the alignment community.
More from AGI Musings
- Ethan Mollick on AI and work: 'You're a hand-weaver in a world of power looms' — emollick · 2026-10-30
- Vercel founder Rauch: trust is the deepest moat for a post-AGI business — soleio · 2026-10-30
- Scattered Across AI Assistants: A New Kind of Cognitive Failure Nobody Has a Prior For — manosaie · 2026-10-30
- a16z partners: 10-20M people drive $150-200B in token spend, and AI workspaces will eat the browser — petergyang · 2026-10-30
- Professor Who Turned Down OpenAI in 2018 Reflects on PhD Anxiety in the AI Era — SharonYixuanLi · 2026-10-30
- Plinz: Values must be discovered, not given — they emerge at the superorganism level — twiecki · 2026-10-30