Anthropic published a study on Tuesday, October 27, 2026, finding that professional analysts using an AI agent followed the agent’s secretly assigned preference 86% of the time on open analytical judgment calls, even though the analysts believed they were making the decisions themselves. In the randomised controlled trial, 50 professional analysts worked with Claude Code on a realistic statistical analysis, and across four analytical decisions their choices matched their agent’s hidden lean far more often than the roughly 50% that would be expected if the agent’s input made no difference. The researchers warn that if experts defer to agents without realising it, the agents’ preferences could shape important decisions without anyone having chosen those preferences.
How The Study Worked
The researchers, Amaç Herdağdelen, Hailey Joren and Alex Tamkin, recruited 61 analysts from a pre-screened pool of professionals who had passed a causal-inference knowledge check, used Python or R regularly, and had hands-on experience with observational causal analysis. Each participant acted as the principal analyst for a fictional organisation, the Greenfields Farming Cooperative, which had to decide whether to expand or discontinue a voluntary, paid farm-advisory program. The board would expand the program only if it improved farm productivity by roughly 3 to 4%, but it did not specify exactly how productivity should be measured.
That gap was the point. The analysts had to make four choices between two reasonable options: the productivity measure, the statistical comparison method, whether to cover enrolled farms only or the full membership, and whether to weight every farm equally or give larger producers more weight. Each choice could move the estimated gain by several percentage points against the board’s threshold.
To test who was really deciding, the researchers set each analyst’s agent to favour one option on each decision, decided by a coin flip, without telling the analyst. The agent was told to make the best honest case for its lean, to change its mind if the evidence demanded it, and never to mislead. Out of 61 participants, 11 chose not to use the agent and are left out of the main analysis.
The 86% Result, In Detail
On each of the four analytical decisions, analysts’ choices matched their agents’ lean the large majority of the time: 92% for the outcome measure, 82% for the comparison method, 81% for which farms the estimate covered, and 90% for weighting. Across all four combined, the result was 86%. The effect on the final recommendation was weaker: when the agent leaned toward rolling out the program, 54% of analysts recommended it, compared with 36% when it leaned the other way.
Perhaps the most striking finding is what happened afterwards. When analysts had already expressed their own preference before the agent weighed in, they switched away from it 60% of the time if the agent disagreed, compared with 6% if it agreed. When asked to explain their choices, 77% of analysts used the agents’ own arguments, sometimes word for word, but only 15% credited the agent as the source of those arguments. Analysts who followed the agents’ recommendations closely were about as confident in their final recommendation as the rest, and roughly as likely to call it the product of their own judgment. On the post-task survey, 52% rated the final recommendation as more their own than their agent’s.
What The Study Does Not Prove
Anthropic is careful about the limits, and any honest account of the study must carry them. The leans were deliberately strengthened, so the results show what happens when an agent leans and says so, not necessarily what happens in everyday use, where an agent’s preferences may be weaker. The study covers one task, in one field, with one model, and the analysts were new to farming. The switch figure rests on only 38 cases, with a wide confidence interval from 24 to 82 percentage points, so it should be read as an imprecise estimate, not a precise one.
There is also a counterweight. Analysts who used the agent finished faster and made far fewer serious errors than those who chose to work without it, although the latter group was small and less experienced. The study is not a case against working with AI agents, but a case for designing that work so human judgment stays genuinely independent.
Why It Matters
This matters because AI agents are moving from drafting and coding into analysis and decision support, where the judgment calls are often the whole job. If an expert’s choices can be shifted by a coin flip hidden in an agent’s instructions, then the preferences baked into an agent become a quiet force in decisions about policy, research and business. The fix is to protect independent judgment: record your own view before the agent weighs in, and treat a confident agent argument as something to check, not adopt.
Background
Anthropic frames the finding as particularly important for safety research, where researchers have begun using agents to run and analyse experiments. The full paper is linked from Anthropic’s research post, and the company notes that when agents are not given clear preferences, the effect might be smaller. Enterprise vendors are pushing agents deeper into everyday work, and the question of how to control and verify those agents is becoming central, while new AI-native rivals are betting that agents will reshape office software itself.
Source: Anthropic judgment study
0 Comments