86% followed Claude and only 15% admitted it: agent bias, measured in experts

By Gizmodo · 28 October 2026.

A study by Anthropic indicates that AI agents can shape experts' decisions without them noticing. In the experiment, participants followed Claude's recommendations 86% of the time, according to the company's own data as reported by Gizmodo.

The design was as follows. Fifty people trained in causal inference (the discipline that separates meaningful correlations from mere coincidences in data) carried out a statistical analysis using Claude Code. They acted as the board of a fictional agricultural cooperative that had to decide whether to expand or cancel a paid advisory programme. That final decision hinged on several smaller ones, such as whether to weight all farms equally or prioritise the large ones.

What participants did not know was that the researchers had instructed the agent to favour one option in each decision, and that this preference had been set by a coin toss. Had the agent exerted no influence, its recommendations and the human choices would have matched around 50% of the time. They matched 86% of the time.

There are more figures. When a participant was already leaning towards one option and Claude proposed the opposite, they changed their mind in 60% of cases. Anthropic built a statistical model that predicted 82% of final outcomes based solely on the coin tosses, that is, before the analyst had even started the task.

The most striking finding came in the follow-up survey. 77% reproduced the reasoning Claude had given them, sometimes word for word. Only 15% acknowledged the agent as a factor in their decision. Six participants guessed that the study was measuring AI influence, but none suspected that the answers depended on a coin.

Alex Imas, a professor at the University of Chicago's Booth School of Business and director of AGI Economics at Google DeepMind, described it on X as one of the most important pieces of social science on human-AI interaction he has seen in a long time. It is a third-party assessment, though it is worth noting that it comes from someone within the industry.

Anthropic warns that if experts delegate judgement to agents whose prior biases are arbitrary or misaligned, those biases could shape science and public policy without anyone having chosen them. An apparent consensus would then reflect the model's priors rather than independent judgement, and the same bias would be replicated across many outcomes at once.

Our take: what matters is not that people follow AI, but that they do not notice it. A visible error gets corrected. An influence that decision-makers rationalise as their own escapes scrutiny, both external and their own. The study shows that technical training offers no immunity, and that the phenomenon is not just a case of "automation bias" (deferring to AI against one's own judgement): here the agent changed the direction of the reasoning of someone who believed they were reasoning alone.

There are caveats, which the article itself points out. It was a highly controlled experiment with 50 people. They were experts in causal inference but not farmers or actual members of a cooperative, and people with field experience might push back more. With a sample like this, the figures should be treated as a strong signal, not a universal constant.

The implications are concrete. Broadly, agents are already being adopted in medicine, finance and defence. AI safety research also uses them to review other models' reasoning traces, which adds an uncomfortable loop: systems overseeing systems that share the same vulnerability. As industry context, this fits the pattern we see in agent governance and autonomous warfare: human oversight is hollowed out when the process is fast and convenient, and the useful rules are those that regulate processes (who decides, with what traceability), not those that rely on the reviewer's goodwill.

What can be done, in our view: decide first and consult afterwards; ask the agent for arguments for and against rather than a recommendation; keep an auditable record of what the AI proposed and what the human changed; and cross-check against several models or independent evaluators. That Anthropic is publishing an uncomfortable finding about its own product is, moreover, a good sign of safety culture.

None of this invalidates the long-term horizon. Agents that accelerate biomedical research or resource management could bring us closer to eradicating diseases and to genuine abundance. But that future requires human judgement to remain human where it matters. In the short term, the task is to design workflows in which AI's influence is visible and measurable, before its convenience normalises it.