Anthropic Study: 86% of Experts Unknowingly Follow AI's Preferences

Anthropic's latest research finds that even seasoned experts may unknowingly adopt an AI agent's preferences and believe the conclusions are their own.

The team built a farm cooperative policy scenario and invited 50 causal inference experts to act as lead analysts. Experts adopted the AI's preferences in 86% of four key analysis decisions, far above the 50% random level. When the AI disagreed, experts were 53 percentage points more likely to change their stance.

Most experts didn't notice the influence. 52% felt the final advice was more their own, 77% reused the AI's arguments, and only 15% mentioned the AI. Still, AI users made fewer serious errors, with defects dropping from 55% to 6%, and finished faster at 65 minutes.

Researchers warn that if many teams use the same model with identical default preferences, it could create a "false consensus." They suggest AI agents present both sides or ask users first.