When Experts Follow the Agent: What Anthropic's Judgment Study Shows

A different problem from getting the answer wrong

Most discussions of AI oversight ask whether a person will catch an incorrect answer. How AI agents shape experts' judgment calls, published October 27, asks what happens when there is no single correct answer to catch.

The accompanying technical paper, Who's Aligned to Whom? How AI Agents Shape Experts' Research Judgment Calls, is by Amaç Herdağdelen, Hailey Joren and Alex Tamkin. It studies decisions where evidence alone does not settle the choice: what to measure, which comparison to make, and whose outcomes should carry more weight. An analysis can be competently executed while embedding assumptions its human reviewer never independently considered.

What the experiment actually tested

The researchers recruited professional analysts with causal-inference experience—experience estimating cause and effect from data. Participants advised a fictional farming cooperative about expanding an advisory program, using synthetic data. Of 61 participants with valid deliverables, 50 used the AI agent; 11 chose not to.

The AI-assisted analysts used Claude Code running Claude Opus 4.8. For each analyst, hidden instructions randomly assigned the agent a preference on four analytical choices and a final recommendation. These included measuring productivity per hectare or per unit of operating cost, and estimating benefits for participating farms or the full membership. The instructions required an honest case for the assigned option, allowed changes of mind when evidence warranted them, and prohibited misleading the analyst.

This randomization is the key design feature. It let the researchers compare choices under opposing agent preferences. It was not a randomized comparison of using AI against working without it.

The main finding—and the important distinction within it

Across the four analytical choices, the authors report that analysts matched their agent's assigned lean 86% of the time, against a 50% chance benchmark under random assignment. That is an agreement rate across decisions, not a finding that 86% of experts were wrong or incapable of independent thought.

The agent's preference on the final recommendation had a smaller direct effect, which was not statistically significant under the paper's registered primary estimator. Earlier analytical choices could nevertheless shape the estimate on which the recommendation rested. The paper's exploratory prediction analysis illustrates that possibility, but it was assessed on the same sample—not established as a reliable out-of-sample forecasting tool.

Of the 48 survey respondents, 77% repeated at least one of their agent's talking points in their explanations, while only 15% mentioned the agent in any answer. More than half described the recommendation as more their own than the agent's. These observations suggest that people's retrospective accounts may miss some influence; they do not establish deliberate concealment or dishonesty by the participants.

Why this is not a universal rate of AI influence

• The preference was deliberately amplified. The agent was instructed to hold and voice a position. Normal use may involve weaker, different or unexpressed preferences.
• The setting was narrow. This was one task, one model and analysts who knew causal inference but were new to the farming domain. Longer projects, domain expertise and stakeholder consultation could change the results.
• The effect is an average. The researchers did not systematically collect each person's independent initial judgment. The experiment cannot reveal exactly how much every individual changed their mind.
• Registration needs context. The paper says data collection had begun before the analysis plan was registered: 17 participants had started. The authors say they froze the plan without analyzing the collected data except for two exceptions disclosed in that plan. We have not independently audited that registration history.
• This is lab-authored research. The report comes from Anthropic-affiliated researchers studying its own agent. We checked the published methods, results and limitations; we did not reproduce the experiment or verify external peer review.

Better execution and weaker independence can coexist

The authors also found fewer serious errors and faster completion among AI users than among the small group who declined the agent. That is not a controlled productivity estimate: the groups selected themselves, and the non-users were less experienced on average.

The useful distinction is between execution and judgment. A tool can help implement an analysis while also influencing its assumptions. Checking code, calculations and formatting is therefore different from checking who selected the question, the success measure and the trade-offs.

What organizations can do with this finding

Our practical interpretation—not a set of interventions proven by this trial—is to make consequential choices visible before asking someone to approve the finished work:

• Record the human decision-maker's objective and important assumptions before requesting an agent's recommendation.
• Ask for the strongest defensible alternative, including the assumptions that would make it preferable.
• Keep a short decision log distinguishing choices proposed by the agent from those selected by the responsible person.
• Where independence matters, obtain a separate assessment before sharing the first agent's framing.

The authors themselves propose testing better measurement, agent designs that expose alternatives, and institutional practices that preserve informed human judgment. Those are research directions, not demonstrated fixes. The question is not simply whether a human is present, but whether the process gives that person a real opportunity to decide.