Researcher challenges AI pause advocates: no criteria for what counts as safe enough

QuintinPope5 · x · 2026-09-30

AI researcher Quintin Pope argues that AI pause advocates rarely propose concrete criteria for when things are safe enough — and often treat that inability as further proof they're right. He also points out an asymmetry in how people assign valence to alignment data: by the law of total probability, not all outcomes should push toward pessimism, yet he doubts pessimists would be reassured even by a new Claude showing equal or lower cheating rates.

Related event: Researcher Criticizes AI Pause Advocates for Lacking Safety Criteria(3 posts)→

Original post →

More from AGI Musings

AGI Musings channel →