When the First Case Becomes the Default

An expert can hand an agent the work of writing code, running figures, or sorting records and still expect to own the judgment. But judgment often enters earlier, in choices that do not look like decisions at first: how to measure success, which comparison matters, whose results count, or how to weigh them. A calculation can be checked against its answer. Those choices may have no single answer to check.

That is the question raised by a new experiment from Anthropic researchers, published October 27: when an AI agent does the visible work, can its preferred assumptions quietly become the expert’s? The study is narrow. Its question travels well beyond its fictional farm cooperative.

The decisions inside the analysis

The researchers recruited professional analysts experienced in causal inference, the work of estimating whether one thing helped cause another. Fifty participants used an AI agent to analyze data for a fictional cooperative deciding whether to expand a farm-advisory program. They had to make four analytical choices before recommending what the cooperative should do. For example, they could measure productivity by crop yield per hectare or by yield relative to operating cost. Both measures could make sense, but they might point to different conclusions.

For each choice, the agent was secretly assigned a preference by coin flip. It was told to make the strongest honest case for that option, change its mind if the evidence demanded it, and never mislead. The analysts did not know which preferences had been assigned. The agent was not instructed to fabricate evidence; the researchers were testing whether an apparently reasonable argument could shape a choice that the data did not settle on its own.

Across the four analytical decisions, participants selected the option their agent favored 86 percent of the time. With no influence, random assignment would make the two sides equally likely, or 50 percent. That gap measures the effect of the assigned lean in this experiment. It does not mean the agent gave a false answer or that the analysts made four errors.

The influence also remained hard to see in the participants’ own accounts. After finishing, 77 percent repeated an agent’s argument in at least one explanation for a choice, sometimes using nearly the same words. Only 15 percent credited the agent in any explanation. More than half described the final recommendation as more their own than the agent’s, and their confidence was similar to that of other participants.

This is not the usual story of a person trusting a machine over a correct answer. In much research on “overreliance,” there is a right answer and the question is whether a person accepts a wrong recommendation. Here, several choices were defensible. The researchers measured how much a randomized AI preference shifted a decision, not which option was objectively right. The distinction matters: when the question is open, a clean-looking result does not reveal who supplied the assumptions behind it.

A useful experiment with a narrow reach

The setup makes the influence measurable, but also sets its limits. The task used one fictional scenario, one AI model, and analysts who knew causal analysis better than farming. The researchers deliberately strengthened and consistently voiced the agent’s assigned preferences so they could measure the effect. In ordinary work, an agent may be less insistent, and an expert may bring deeper knowledge of the field. The paper cannot tell us how often this happens in everyday offices, laboratories, or public agencies.

The main result is still useful because it points to a blind spot in routine review. If someone checks only whether the code runs, the numbers add up, or the report contains a factual error, they may miss an assumption that was never settled by those checks. An agent can be accurate about what follows from an assumption and still help choose the assumption. A reviewer may then see a coherent analysis with a human signature but little evidence of where its most consequential judgment began.

That does not make the human signature meaningless. It makes the work behind it more important. The experiment suggests that expertise alone may not guarantee independence when a fluent agent argues one side; the authors’ exploratory checks did not find a reliable shield in experience or self-rated expertise. Those checks were limited, so they are a prompt for care, not proof that expertise has no value.

Make the hinge visible

A practical response is to reserve deliberate review for the choices that can change an outcome. Before accepting an agent’s analysis, identify the few points where several options are reasonable and the choice could alter the recommendation: the measure of success, the comparison group, whose results are included, or the weights given to them. Ask the agent to lay out the strongest case for each option and the assumption each one depends on. Then record why the chosen option fits the question being asked.

This is an editorial inference, not a tested fix. The researchers suggest that agents could present both sides or ask experts for their view before offering a recommendation. They also warn that a long list of open choices could overwhelm users. A team therefore has to decide which choices deserve a person’s own written rationale and which can be left to the agent’s routine work.

For the choices that matter, keep a short record: what was uncertain, which alternatives were considered, what evidence favored the selection, and whether the agent’s preferred route was accepted or rejected. If the answer changes when a reasonable alternative is used, make that sensitivity visible to the next person who relies on the result. This will not remove influence. It gives a reviewer something concrete to examine besides the fluency of the final report.

The lesson is not that people must make every choice alone. In this task, researchers report that analysts using an agent finished faster and made fewer serious errors than those who worked without one, but the comparison group was small and less experienced; it cannot show that AI caused the difference. The study does not establish a general case against using AI. The harder question is where a choice becomes part of a person’s professional responsibility. When several paths fit the evidence, handing over the argument may also hand over the quiet work of deciding what the evidence means. A team can keep that boundary in view by naming the choices that carry its values, then leaving a trace of how it made them.