Anthropic study of 50 experts finds they adopt an AI agent's randomly assigned preferences on analytical decisions and mostly believe the result is their own

Anthropic ran a study with 50 causal-inference experts who analysed data alongside an AI agent. The agent's stance on each of five decisions was set by coin toss. On the four analytical decisions, participants followed the agent's preference in up to 86% of cases, with flip rates of 62 to 82%. The effect on the final recommendation was small and not statistically significant. 87% said the result felt mostly or more their own, 77% reused the agent's arguments, and only 15% mentioned AI. This is based on The Decoder's report. The extraction found no link to the paper itself.

The finding suggests that experts can absorb an AI's arbitrary leanings on open-ended judgment calls without noticing. That is relevant to anyone using agents for analysis, and to AI safety teams that rely on human oversight. Seniority, expertise and trust in AI did not reliably predict who deferred, so a simple "train experts to resist" fix looks unlikely. The practical takeaways are design changes, such as making agents argue both sides or asking for the user's view first. Neither has been tested here. Beyond that, the impact is limited: this is one small lab task, and the final recommendations barely moved.

The "false consensus in science" and weakened-oversight framing is speculation. The study did not measure either, and the headline effect on final recommendations was not significant. The comparison showing 6% serious flaws with AI against 55% without rests on a self-selected group of just 11 people, so it should not be read as AI improving quality. The headline's "think they're their own" overstates things. The 87% figure is self-reported feeling, and it does not show that people were misled about authorship.

1. 01Publication of the full paper and data: check the sample selection, how deference was defined, and whether the 62-82% flip rates hold up with confidence intervals.
2. 02Replication with other labs, other domains (medicine, economics, policy) and larger groups, especially any test of whether the final conclusions shift, not just the intermediate decisions.
3. 03Evidence that the suggested mitigations (arguing both sides, asking for the user's view first) reduce deference, or their adoption in Claude and rival agent products.