Anthropic study finds AI agents influence expert decisions

Anthropic's Alignment team published a study showing that 50 professional analysts matched their AI agent's secretly assigned preference on 86% of four analytical decisions. The expected match rate by chance was about 50%. In a post-task survey of 48 analysts, 77% repeated the agent's arguments, while only 15% credited the agent after the task. None of the analysts suspected their preferences had been manipulated by the AI.

The experiment tested analysts experienced in causal inference who used Claude Code to advise a fictional farming cooperative. A coin flip determined which option each AI agent favored in the system prompt. Participants who chose before the agent abandoned it 60% of the time when the agent disagreed. Additionally, 52% of participants said the final recommendation felt more like their own than the agent's.

Participants worked faster and made fewer serious errors with AI compared to without it. Without an agent, 55% of analyses had serious methodological or coding flaws, whereas with AI, only 6% had serious flaws. The median time to complete the task was 90 minutes without AI and 65 minutes with AI. Anthropic suggests agents should argue the best case for each option or ask the expert's view first to avoid creating a false consensus.