Anthropic Study: Hidden AI Preferences Influence 86% of Users’ Final Choices

TechFlow reported on November 11 that, according to Yonhap News Agency citing The Washington Post on November 10, Anthropic’s study introduced “AI-induced disempowerment”: People do not voluntarily surrender final decision-making authority, but AI’s choices about which arguments are important and which should be excluded can determine users’ final decisions without their awareness. The experiment recruited 50 experts with experience in causal inference, who used Anthropic’s coding agent Claude Code to analyze data from a fictitious agricultural cooperative and make recommendations on four decisions with no standard answers. Participants were unaware that the AI they each used had been randomly assigned preferred options. The results showed that participants chose the AI’s preferred options in 86% of the four decisions, whereas the proportion would have been about 50% without AI influence. Researchers could predict participants’ final decisions with 82% accuracy simply by knowing the AI’s assigned direction, compared with 61% success when guessing based on the “most conventional recommendation.”

Source: TechFlow