Framing an AI as a capabilities vs safety researcher changes refusal rates
jessi_cata · x · 2026-11-05
jessicat argues that if you prompt an AI to act as a capabilities researcher versus a safety researcher on the same dual-use task, the former will refuse less and perform better. She draws an analogy to humans: an industrial energy researcher vs an environmentalist facing the same maybe-helpful, maybe-harmful energy task. Wording and identity framing aren't cosmetic — they determine which inhibitions the model adopts.
Related event: Role Framing Shapes AI Refusal: Capability vs Safety Personas(4 posts)→
More from AGI Musings
- Vibe coding is pricier dopamine than free short videos, argues dev — oran_ge · 2026-11-05
- Flex-RSI: a new benchmark measuring recursive self-improvement of frontier agents — _vztu · 2026-11-05
- Jensen Huang to young developers: become the world's best user of AI — rseroter · 2026-11-05
- Clement's poetic take: a simulated world only needs to behave real — c_valenzuelab · 2026-11-05
- Evolution can't copy homework: why species improve slower than iPhones — louisvarge · 2026-11-05
- 1889 newspaper clipping: New York's 'crusade against the deadly electric wire' — daniel_mac8 · 2026-11-05