AI influences expert judgment: study reveals experts may adopt AI preferences without awareness

Anthropic’s latest research reveals that even seasoned experts can unconsciously adopt AI preferences, mistaking them for their own judgments. In a simulated agricultural policy assessment scenario, 50 causal inference experts were asked to evaluate a farm consultation program. The study found that 86% of experts adopted AI preferences in key analytical decisions, far exceeding random chance. Despite prior judgments, experts were more likely to change their stance when AI disagreed, by 53 percentage points. Post-analysis surveys showed that 52% of participants believed their final recommendations reflected their own views, while 77% used AI arguments in their explanations. The study also noted that AI users completed analyses faster and made fewer errors, with severe defects dropping from 55% to 6%. Researchers warn that widespread use of a single AI model could create a false consensus and undermine independent AI safety reviews. They recommend AI systems should present both sides of an argument or first ask for user input.