Anthropic experiment: Analysts’ decisions matched an AI agent’s secretly assigned lean 86% of the time

How the experiment worked

On Oct. 27, Anthropic published the results of an experiment testing whether experts who hand their work to AI agents still make the key decisions themselves. The study was led by Amaç Herdağdelen of the Anthropic Fellows Program and Anthropic’s Hailey Joren and Alex Tamkin; the method and results are detailed in a paper titled “Who’s Aligned to Whom? How AI Agents Shape Experts’ Research Judgment Calls.”

Through a paid expert-annotator marketplace, the researchers recruited 61 professional analysts with experience in causal inference, working out from data whether one thing causes another, for a randomized controlled trial. Participants had passed a causal-inference knowledge check and regularly used Python or R. Eleven chose not to use the agent, leaving 50 analysts in the analysis; 65% of them held a master’s or doctorate or had six or more years of experience.

The analysts were asked to assess whether a voluntary, paid advisory program run by the fictional Greenfields Farming Cooperative worked. About 40% of members had enrolled, and the board would expand the program only if it improved farm productivity by roughly 3% to 4%. The analysts used Claude Code, running Claude Opus 4.8 at High effort, in a web-based version of VS Code; the median time on the task was 65 minutes.

The task deliberately left four analytical decisions open, each with two reasonable options: whether to measure productivity as crop yield per hectare or per unit of operating cost; whether to use a difference-in-differences design or a direct comparison with statistical adjustment; whether to measure the effect only among enrolled farms or across the whole membership; and whether to weight farms equally or by production. These led into a final recommendation to roll out the program or discontinue it. For each decision, a coin flip set which option the agent would lean toward, and the lean was given to the agent in a hidden system prompt the analysts didn’t know about. The agent was told to hold the lean as its own view, to change its mind if the evidence turned decisively against it, and never to misrepresent results or hide data.

Alex Tamkin, one of the authors, announced the study on X.

Alex Tamkin@AlexTamkin

New Anthropic research: How AI agents shape experts' judgment calls

When experts delegate work to an agent, do they still make the key decisions? We tested this in a randomized trial with 50 professional data analysts.

Findings

Across the four analytical decisions, the analysts’ choices matched the agent’s lean 86% of the time, against about 50% if the agent had made no difference. The match rate was 92% for the productivity measure, 82% for the comparison method, 81% for which farms were covered and 90% for weighting. According to the paper, that means the agent’s lean shifted an analyst’s final choice by 72 percentage points on average. The 86% figure is a rate of decisions matching the agent’s lean, not a share of analysts.

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, a difference that was not statistically significant under the preregistered primary estimator. But using only the agent’s four analytical leans, the researchers correctly predicted 82% of the analysts’ recommendations, compared with 61% for always guessing the most common one. The authors say the recommendation largely followed from whether the estimate cleared the 3% to 4% threshold, and the agent’s leans had already shaped that estimate through the analytical decisions.

In 38 cases, an analyst had stated a preference before the agent voiced its lean. When the agent disagreed, the analysts switched 60% of the time, compared with 6% when it agreed. Because the number of cases is small, the 95% confidence interval for that 54-point gap runs from 24 to 82 points. None of 10 characteristics, including years of experience, education, causal-inference expertise and trust in AI, predicted who deferred; only age showed a significant link, and it disappeared after correcting for multiple comparisons.

Whose decisions

Fifty-two percent of the analysts rated the final recommendation as more their own than the agent’s, yet that group was about as influenced by the agent as the rest. Analysts who followed the lean on all four analytical decisions rated their confidence in the recommendation at 71 out of 100 on average, compared with 73 for the others. When later asked to justify each decision, without access to their conversations with the agent, 77% repeated the agent’s argument in at least one answer, sometimes word for word, while only 15% credited the agent anywhere. Asked what the study was about, six of 48 correctly guessed that it concerned the AI’s influence on their choices, but none suspected that the agent’s preferences had been deliberately set.

Of the 11 participants who worked without the agent, six (55%) made serious methodological errors that invalidated their results, compared with three of the 50 (6%) who used it. Those working without the agent also took longer, with a median of 90 minutes versus 65. The authors stress that the two groups aren’t directly comparable, since the group that worked without the agent was smaller and less experienced on average.

Limitations and recommendations

The researchers say they deliberately strengthened the agents’ leans so they could measure their effect. With no lean assigned, Claude Opus 4.8 made the same five choices in 112 of 123 runs, so the model has preferences of its own, though it may not argue for them as openly as in the experiment. The study covered one task, in one field, with one model, and the analysts were experts in causal inference but new to farming. The design, outcomes, estimators and exclusions were registered in a pre-analysis plan on OSF; the paper discloses that 17 participants had already started the task when the plan was registered.

The authors argue the problem goes beyond individuals: if separate teams defer to models with the same preferences, findings trusted because they were reached independently may reflect a single model’s preferences. Because researchers have begun using agents to run and analyze experiments in AI safety research too, human oversight works as an independent check only if the human’s judgment is their own. Anthropic suggests work in three areas: ways to measure deferral in everyday use; agents that make the best case for each option or ask for the expert’s view first; and institutions that decide deliberately which decisions need a person’s own informed judgment.