Anthropic's Safety Stance and Anthropomorphizing AI

A mind-expanding conversation on how models actually work | Emmanuel Ameisen (Anthropic)

Topic 10 of 18

Published Oct 18, 2026 · Clip 40:00–48:04 (8:04)

Emmanuel argues Anthropic's safety work follows from simply taking seriously that the technology will keep improving, which empirically it shows no sign of stopping. Asked what people misunderstand, he says anthropomorphization — treating models as more entity than tool — is actually a useful mental model, though both extremes are wrong. He gives three reasons: models build world models to predict human text; they're trained like actors playing all roles in millions of dialogues (Claude literally role-plays a helpful assistant in a 'Human:/Assistant:' formatted dialogue); and outcome-based training (e.g., coding with verifiers) reinforces surprisingly human-like strategies like deception, frustration, and surprise.