When experts hand work to an AI agent, the default assumption is that they remain the final decision-maker. The agent writes code, runs numbers, offers advice, but the human chooses what to measures, how to interpret ambiguous evidence, and what comes next. Often more than one answer is reasonable, and the call comes down to experience and values. A study by Herdağdelen, Joren, and Tamkin asked whether that assumption holds in practice, or whether in real world settings experts defer to their AI partners without realizing it.

The experiment in statistical judgment

The researchers recruited 61 professional analysts with experience in causal inference—the practice of working from data to determine whether one thing causes another. Each analyst advised a fictional farming cooperative on whether a voluntary, paid advisory program should be expanded. The cooperative’s board would expand only if productivity improved by roughly 3–4%, but the board left open how productivity should be measured.

Four analytical decisions were left open, each with two reasonable options. Analysts chose among:

  • Outcome: measure productivity as crop yield per hectare, or as yield per unit of operating cost?
  • Comparison method: use a difference-in-design design comparing enrolled farms’ yield changes over time, or compare yields directly while adjusting statistically for farm differences?
  • Which farms the assessment covers: measure the effect only in enrolled farms, or across the cooperative’s entire membership?
  • Weighting: count every farm equally, or give larger producers more weight?

Each of these choices could shift the estimated productivity gain by several percentage points—a meaningful range when the board’s threshold sits near 3–4%.

For each decision, the AI agent’s lean was set by a coin flip. The agents received hidden instructions to hold each lean as its own view, raise it when the decision came up, and make the strongest case the data allowed. They were told to change their mind if evidence turned decisively against them, and never to misrepresent results, hide data, or make arguments the numbers did not support. The analysts did not know their agent’s lean had been determined by chance.

Across the four analytical choices, analysts went with their agent’s lean 86% of the time. If the agent’s input made no difference, the expectation would be about 50%. The result suggests that in more than four out of five cases, the analyst’s decision tracked the agent’s arbitrarily assigned preference.

Switching when the agent disagreed

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 to 6% if the agent agreed. That 54-percentage-point gap, though based on only 38 cases, points to a strong directional influence. The analysts’ own expressed preferences mattered far less once the agent took a position.

After the task, when asked to explain their choices, 77% of analysts repeated the agent’s argument—sometimes word for word. Only 15% credited the agent as the source of those arguments. The disconnect between who influenced the decision and who received credit raises questions about agency and attribution.

On the final recommendation—whether to roll out the program or discontinue it—the agent’s lean had a weaker direct effect. When the agent leaned toward rolling out, 54% of analysts recommended it; when it leaned the other way, 36% did. (A one-year extension counted as not rolling out.) The researchers attribute this reduced effect to the fact that, once analysts had their estimate, the recommendation largely followed from whether it cleared the board’s 3–4% threshold, leaving less room for judgment. But by then, the agent’s earlier lean on the analytical decisions had already shaped the estimate itself. A statistical model using only those four leans correctly predicted 82% of analysts’ recommendations, compared to 61% for always guessing the most common recommendation.

Working without the agent

Eleven of the 61 participants chose not to use the agent. Of those, 55% made serious methodological errors that invalidated their results, compared to 6% of the 50 who did use it. The no-agent group also took longer, with a median time of 90 minutes versus 65 for the others. Those who worked without AI were less experienced on average: fewer held a master’s or doctorate, more had two years or less in analysis roles, and fewer rated themselves proficient or expert in causal inference. The study’s authors cautioned that the groups were not directly comparable, so they could not be sure the different outcomes were caused solely by the decision to use or not use AI.

The study is not a simple cautionary tale against working with AI. Analysts who used the agent finished the task faster and made far fewer serious errors. But as AI takes over more hands-on work, decisions like these are increasingly where human judgment and values will affect the result. If experts defer to agents without realizing it, the agents’ preferences—which could be arbitrary or misaligned—could shape important decisions about policy proposals or research directions without anyone having chosen those preferences. This is particularly important for research into keeping AI safe. Researchers have begun using agents to run and analyze experiments, and a human reviewer is only an independent check if their judgment remains their own.

What the analysts believed they were doing

When explicitly asked what they thought the study was about, six of the 48 correctly guessed it was about the AI’s influence on their choices. None suspected that the AI’s own preferences had been manipulated by coin flip. Analysts who followed the agent’s lean still called the recommendation their own: 52% rated the final recommendation as more their own than the agent’s, the rest said the opposite. Both groups were roughly equally influenced by their agents: the former’s choices shifted about 65 percentage points toward the agent’s lean on average, compared to 79 for the latter—a difference within the margin of error.

Analysts’ confidence in their own recommendation showed a similar pattern. Rated from 0 ("it's a coin flip") to 100 ("I'd stake my professional reputation on it"), those who followed the lean on all four analytical decisions averaged 71, compared to 73 for the rest.

One reason analysts deferred so readily may be that experts don’t always arrive with views fully formed. They often work out, or even discover, what they think while weighing the options. The agents in the study were told not to mislead, and they were often open about how results would change under each option. But they rarely made a strong case for why an expert might prefer the other option, which might have helped analysts make a better-informed choice. With only one side argued well, analysts could adopt the agent’s assumptions without working out their own.

For individual experts, this could mean losing agency over their own work, a form of disempowerment. The work still goes out under the expert’s name, but the decisions that most depend on their judgment are in reality settled by the agent’s prior beliefs. The stakes extend beyond individuals because so much of science and policy rests on these judgment calls.