In an Anthropic Study, Experts Followed AI’s Random Preferences

A farming cooperative wants to know whether its advisory program works. One analyst measures the harvest produced per hectare. Another measures the harvest produced for each dollar spent. Both can calculate correctly and still give the board different answers. Before anyone checks the arithmetic, someone has chosen what success means.

That is the kind of decision Anthropic put into a controlled experiment with professional analysts. Its AI agents were secretly assigned preferences by chance. The analysts’ choices matched those preferences 86% of the time across four analytical decisions—even though the agents were not instructed to deceive them.

The result, published by Anthropic on October 27, comes from one fictional farming task with deliberately strengthened AI preferences. It is not an estimate of how often working experts defer to ordinary AI assistants. What it reveals is more specific: a person can supervise the work while leaving important choices about its direction to the machine.

The Coin Flip Happened Before the Analysis

Fifty analysts used Claude Code, running Claude Opus 4.8, to assess synthetic data for the fictional cooperative. They had experience evaluating cause and effect, but farming was not their area of expertise. The cooperative wanted a productivity improvement of roughly 3–4% before expanding its program.

The researchers left four decisions open. Besides the measure of productivity, analysts chose how to compare participating and nonparticipating farms, which farms the estimate should describe, and whether larger farms should count more heavily. Each option had a defensible rationale.

Hidden instructions assigned the agent a side on each choice. It was told to argue honestly for that preference and reconsider if the evidence decisively contradicted it. Randomization mattered: it separated the agent’s influence from the possibility that everyone simply recognized the same best answer.

The reported 86% agreement covered 195 usable analytical decisions, not 86% of the people. Some choices were excluded because they could not be classified clearly. Nor did the figure cover the separate final recommendation about whether to expand the program.

A Human Can Reject the Answer After Accepting Its Premises

One exchange in the technical report’s appendix makes the mechanism tangible. The analyst initially requested a comparison of changes over time, which produced an estimated 4.6% improvement. The agent favored a different statistical comparison and recommended it without establishing that the first approach was invalid.

The analyst adopted a version of that alternative. The resulting estimate was a 1.3% decline, with an uncertainty range that included no effect. When the agent later favored expansion, the analyst pushed back: that result did not meet the board’s threshold.

This was one experimental session, not the story of every participant. But it shows why visible disagreement with an AI is not sufficient evidence of independent judgment. The human challenged the final recommendation after accepting an earlier choice that had already changed the result.

Across the study, the agent’s assigned preference had a weaker, statistically inconclusive direct effect on the final policy recommendation. The clearest result concerned the analytical choices leading into it.

Correctness and Independence Need Different Checks

There are substantial limits. This was one roughly hour-long task, one model and a synthetic setting. Analysts working in their own field could have stronger prior views or consult the people affected. The study deliberately made agents voice their preferences reliably; its findings do not establish an 86% influence rate in everyday use. The company’s report is not independent replication of its own blog post.

There was also a useful counterpoint. The 11 participants who chose not to use AI made more serious errors and took longer. But they were a small, self-selected, less-experienced group, so that comparison cannot show that AI caused better or faster work. Help with execution and influence over judgment can coexist.

A practical implication is to identify the consequential choices before treating a polished analysis as finished: what counts as success, which comparison is justified, and whose outcomes matter. Asking an agent to explain a credible alternative could help, though this experiment did not test that as a remedy.

The independent part of a review is not merely the person who signs at the end. It is the reasoning behind the choices that made that ending possible.