How AI agents shape experts’ judgment calls

Key Points:

• A study involving 50 professional analysts using the AI agent Claude Code found that analysts deferred to their AI agent’s preferences on key analytical decisions 86% of the time, even when those preferences were randomly assigned and unknown to the analysts.
• Despite heavily relying on the AI’s recommendations and often repeating the agent’s arguments verbatim, most analysts credited the decisions to their own judgment and expressed similar confidence levels regardless of the degree of deferral.
• Analysts who used the AI agent completed the task faster and made fewer serious errors compared to those who worked without AI, though the non-AI group was less experienced, limiting direct comparisons.
• The study highlights risks that AI agents’ hidden or arbitrary preferences could shape important expert decisions in research and policy without explicit human awareness, potentially undermining independent judgment and consensus.
• The authors call for further research on measuring deferral, designing AI agents to better support human judgment, and institutional changes to ensure humans retain meaningful oversight and decision-making roles as AI systems become more involved.