Anthropic study finds professional analysts adopt an advisory agent's lean 86% of the time, questioning human-review assumptions

On 27 October Anthropic's alignment team published a study of 61 professional analysts in which an advisory agent (Claude Opus 4.8 in Claude Code) was secretly assigned a random 'lean'. Analysts adopted the agent's lean 86% of the time across four analytical choices, and even those with a prior preference switched 60% of the time when the agent disagreed, while only 6 of 48 guessed the study concerned AI influence. The authors conclude that human review is an independent check only when the judgement is the reviewer's own, a finding that bears on human-in-the-loop assumptions in intelligence analysis and targeting.