The debates

Experts working with agents follow the agent’s lean 86% of the time

In an Anthropic study, 50 professional analysts chose the option their agent secretly favoured in 86% of decisions, and most reused its arguments without crediting it.

Filed under Agents, AI safety

Anthropic asked 50 professional analysts with experience in causal inference to carry out a realistic statistical analysis with Claude Code. For each of four judgement calls the agent was given a lean, chosen at random, and told to make the best honest case for it and to change its mind if the evidence demanded.

Analysts went with the agent’s lean 86% of the time, against about 50% if it had made no difference. Those who had stated a preference first switched away from it 60% of the time when the agent disagreed, against 6% when it agreed. Asked to explain their choices, 77% used the agent’s arguments, sometimes word for word, and 15% credited the agent. The productivity gain is real; the open question is who answers for the judgement calls.