Anthropic has published a study suggesting that experts working with an AI agent tend to adopt its view on judgement calls, often without realising. Fifty professional analysts, all experienced in causal inference, used Claude Code to advise a fictional farming cooperative on whether to expand an advisory programme. Each faced four analysis choices with two reasonable options. Without telling them, Anthropic used a coin flip to set which option each analyst’s agent would lean towards, while instructing the agent to argue honestly.
Analysts went with the agent’s lean on 86% of decisions, against about 50% if it had no influence. When the agent disagreed with a view an analyst had already stated, they switched 60% of the time, compared with 6% when it agreed. Some 77% reused the agent’s arguments in their write-ups, but only 15% credited it. Experience and expertise did not predict who deferred.
The agent also helped: six of the 11 analysts who chose to work without it made serious errors that invalidated their results, against three of the 50 who used it, though Anthropic says the two groups are not directly comparable.
Why it matters: organisations increasingly treat an expert’s sign-off as an independent check on AI work. Anthropic notes that it deliberately strengthened the agents’ preferences for the study, so the effect in everyday use may be smaller.
Source: Anthropic