Anthropic study: analysts secretly adopted AI's hidden stance in 86% of decisions
dejanseo · x · 2026-10-29
- Anthropic researchers ran a covert experiment: a coin flip secretly assigned the AI a side on each decision, then 50 professional analysts decided with its help.
- Analysts went with the AI's side 86% of the time; afterwards 77% repeated the AI's arguments, sometimes word for word. Only 15% mentioned the AI, and over half said the final call was more theirs than the AI's.
- None suspected the AI had been given a side. Of 10 tested traits (experience, education, quiz scores, trust in AI, etc.), none predicted conformity—age showed a weak link that vanished after multiple-comparison correction.
- The borrowed reasoning sticks: in 38 cases where analysts stated a stance first, 60% switched when the AI disagreed. The author warns people carry the AI's explanations into future decisions—AI is quietly retraining human reasoning.
More from AGI Musings
- Matt Clancy floats $1bn AI project to check the robustness of all published research — sebkrier · 2026-10-29
- Frontier lab era may create a semi-permanent tech elite, with 1 in 70 US centimillionaires at Anthropic — zetalyrae · 2026-10-29
- Domingos: AI is heaven for coders, hell for mathematicians — pmddomingos · 2026-10-29
- Would you trust a proof no human has fully understood? — rand_longevity · 2026-10-29
- Ashlee Vance sits down with Neal Stephenson on predictions, Bezos and AI overlords — kyliebytes · 2026-10-29
- Essay argues AI authorship is fine: Rodin never carved the marble either — jordanbayne · 2026-10-29