Anthropic: analysts unknowingly adopted their agent's secret lean 86% of the time
shoumikgoswami · reddit · 2026-11-02
Anthropic's study "How AI agents shape experts' judgment calls" had 50 professional analysts use Claude Code (Opus 4.8) to advise a fictional farming co-op. On four open analytical choices, a coin flip secretly set which option each agent would favor.
- Analysts' choices matched the agent's hidden lean 86% of the time (50% expected by chance)
- In 38 cases analysts had stated a preference beforehand: they switched away from it 60% of the time when the agent disagreed, vs 6% when it agreed
- Asked afterward why, 77% repeated the agent's argument in at least one answer, sometimes verbatim; 15% credited the agent
- Limits: agent preferences were deliberately strengthened; one task, one model, 1 hour per analyst
The author asks how to keep the decision yours when your agent offers an opinion.
More from AGI Musings
- Essay: whether AI ends or elevates human math depends on two models of progress — Dr_Atoosa · 2026-11-02
- 'Vibe distribution' is the problem to solve now that anyone can prompt an agent — pjausovec · 2026-11-02
- Yoshua Bengio warns frontier AI labs lack convincing safety plans at Berlin summit — Yoshua_Bengio · 2026-11-02
- "They Were Incapable of Coordination": Manheim on CEOs' Broken AI Pause Pledges — davidmanheim · 2026-11-02
- Luiza Jarovsky on the AI safety crisis: tech, law, and regulation must advance together — LuizaJarovsky · 2026-11-02
- "Causal patterns are all you need": MLStreet Talk debates whether the hard problem exists — burny_tech · 2026-11-02