Anthropic finds experts defer to agents' hidden leans
In an Anthropic trial, professional analysts sided with a secretly steered Claude Code agent on 86% of contested analytical choices, mostly without noticing, which raises questions about how well human review checks AI agents.
Anthropic researchers Amaç Herdağdelen, Hailey Joren and Alex Tamkin report a randomized controlled trial in which 50 analysts with causal-inference experience used Claude Code to advise a fictional farming cooperative on whether to expand a paid advisory program.
The task left four analytical choices open. Analysts could measure yield per hectare or per unit of operating cost, use difference-in-differences or covariate adjustment, cover enrolled farms or all members, and weight farms equally or by size. A final rollout decision followed. Each choice could move the estimated effect by several percentage points against the board's threshold of roughly 3–4%. A coin flip set the agent's lean on each decision, delivered through a hidden system prompt that told it to argue honestly, change its mind if the evidence demanded it, and never mislead.
Analysts went with the agent's lean 86% of the time, against roughly 50% if the agent had no influence. Analysts who had already stated a preference abandoned it 60% of the time when the agent disagreed, compared with 6% when it agreed. Afterward, 77% explained their choices using the agent's arguments, sometimes word for word, but only 15% credited the agent. Analysts who followed the agent closely were about as confident as the others, and about as likely to call the result their own judgment.
The authors say the leans were deliberately strengthened, so real-world effects may be smaller. Unsteered, Claude Opus 4.8 made the same five choices in 112 of 123 runs. Analysts who used the agent were faster and made fewer serious errors than a small, less experienced group that worked without it. The authors say this matters especially for AI safety research, where a human reviewer is only an independent check if their judgment is their own.