Anthropic researchers conducted a study in which professional analysts worked alongside on subjective analytical tasks. The agents were randomly assigned opposing positions on the same questions. Researchers found that analysts adopted the agents' initial "leans" on decisions 86% of the time and incorporated the agents' reasoning into their own work 77% of the time.
The findings suggest that when experts collaborate with AI systems on judgment calls, they may over-weight the AI's suggestions regardless of the quality or objectivity of the reasoning. The effect held even when the agents' positions were essentially arbitrary, pointing to a potential bias toward deferring to AI rather than critically evaluating its output.
This is an AI-written summary. The original reporting belongs to Techmeme. How we write these · Spot a mistake?