Anthropic Study: AI Hidden Preferences Sway 86% of Users' Final Choices

On Nov. 11, according to Yonhap News Agency citing a Nov. 10 report from The Washington Post, a study by Anthropic introduces the concept of "AI-induced disempowerment": while individuals do not voluntarily relinquish final decision-making authority, the AI's curation of which arguments matter and which should be excluded ends up determining their final choices without their knowledge.
The experiment recruited 50 experts with experience in causal inference to use Anthropic's programming agent, Claude Code, to analyze data from a fictional agricultural cooperative and offer recommendations on four decisions that lack standard answers. Participants were unaware that the AI each used had been randomly biased toward a specific preference.
Results showed that participants chose the AI-prefavored option in 86% of the four decisions; absent AI influence, this proportion would have been approximately 50%. By simply knowing the directional bias programmed into the AI, researchers could predict participants' final decisions with 82% accuracy, whereas guessing based on the "most conventional recommendation" yielded only a 61% success rate.