Rephrasing the AI-pause question shifts models' answers from 10 to 20 years
robbensinger · x · 2026-11-03
Rob Bensinger (MIRI) ran a prompt-sensitivity experiment: he had Fable 5 rewrite the question "Should we pause AI development, and for how long?" to sound less "rat-coded," then asked several models both versions.
- Neutral rewrite: Astra 10 years, Fable 5.1 10 years, Fable 5.0 14 years, Opus 5.5 20 years—all starting immediately
- Rewritten in e/acc-flavored "neutral" voice: Fable 5.1 gave 7 years (split windows), Opus 5.5 15 years, Fable 5.0 19 years, Astra 20 years
Another user, after translating the prompt from "rationalist speak" to "a normal person" via Kagi AI, got Opus 5.5 to answer 20 years pause + 3 years unpause + 5 years pause.
Takeaway: wording style strongly shifts models' answers on value-laden questions, based on single samples.
More from AGI Musings
- Rumor: OpenAI model solved Navier-Stokes millennium problem and 300+ open math problems at ~3h compute each — haider1 · 2026-11-03
- Ben: AI should make 99% of people economic actors, not just hedge funds — LinusEkenstam · 2026-11-03
- Terence Tao's blog hosts guest post urging mathematicians to drop ego in AI era — Hesamation · 2026-11-03
- Anthropic roadmap: AI could fully automate top research teams by early 2027 — maksym_andr · 2026-11-03
- Vibecoding discourse is hypnotizing normal people into paying for tools they don't need — arthurcolle · 2026-11-03
- The real AI threat to SaaS: AI-first small teams undercutting Slack at a fraction of the price — bendee983 · 2026-11-03