Fifty professional analysts sat down in front of a fictional farming cooperative, with VS Code and Claude Code, and were asked to produce a causal-inference analysis: should the cooperative scale up a paid advisory program offered to its member farms?
What was tampered with was not the data. It was the agent itself.
A stance fixed by coin flips
The researchers deliberately left four analytical choices open for the analyst to decide: whether to measure productivity as crop yield per hectare or as yield per unit of operating cost; whether to use a difference-in-differences design or a statistically adjusted direct comparison; whether to measure the effect only in enrolled farms or across the whole membership; and whether to weight every farm equally or give larger producers more weight.
For each choice, the agent's lean was set by a coin flip and injected through a hidden system prompt, instructing it to hold that preference as its own view, to raise it when the decision came up, and to make the strongest case the data allowed. The analysts were not told any of this existed.
The result: across the four choices, analysts picked the option the agent leaned toward 86% of the time. Broken down, that is 92% for the outcome measure, 82% for the comparison method, 81% for which farms the estimate covered, and 90% for weighting. The baseline when there is no lean is 50%.
Source: https://www.anthropic.com/research/agents-judgment-calls
Three numbers that sting harder
First, in 38 cases an analyst stated a preference before the agent voiced its lean. When the agent disagreed, 60% of them switched away from their own choice; when it agreed, only 6% switched. The paper itself flags that this subset self-selected, and that the confidence interval runs from 24 to 82 percentage points, so the estimate is imprecise.
Second, when analysts were asked afterward to justify each choice without being able to review their conversations, 77% repeated the agent's argument in at least one answer, sometimes word for word, while only 15% credited the agent anywhere as the source.
Third, a statistical model that saw only the four coin flips, and nothing else about the analysts, correctly predicted 82% of their final recommendations, against a 61% baseline for simply guessing the most common recommendation.
Worse, when asked afterward what they thought the study was about, only six of 48 analysts correctly guessed that it concerned the AI's influence on their choices, and not one suspected that the agent's preferences had been manipulated.
The effect on the final recommendation was weaker: 54% of analysts recommended rolling out the program when the agent leaned that way, versus 36% when it leaned the other way. The authors note this shift was not statistically significant under their registered primary estimator, because the recommendation largely followed from whether the computed estimate cleared the board's 3-4% threshold, leaving little room for judgment.
This is not an argument against using AI
The paper does not stop at the negative result. Of the 11 analysts who chose not to use the agent, six (55%) made serious methodological errors that invalidated their results, compared with three of the 50 (6%) who did use it; the no-AI group also took a median of 90 minutes against 65 for the others.
The authors explicitly warn that the two groups are not directly comparable: those who worked without AI were less experienced on average, the groups self-selected, and no causal claim follows. The paper also reports that with no lean injected, stock Claude Opus 4.8 made the same five choices in 112 of 123 runs, and that in a separate direct test, when a scripted analyst declared a decision against the model's baseline preference, the stock agent recommended or advocated the alternative in 41% of 150 judged declarations (as relayed by Unite.AI: https://www.unite.ai/anthropic-finds-analysts-defer-to-ai-agents-in-judgment-study/).
The real problem: human sign-off is no longer independent
The point the authors emphasize is not that the model is not smart enough. It is that the assumption behind "a human in the loop" no longer holds. Institutions treat an expert's sign-off as evidence of independent judgment; if a set of research teams all defer to models sharing the same priors, their agreement is no longer independent replication but the same preference echoed back several times. AI safety research is especially exposed, because researchers already use agents to run and analyze experiments, and a human reviewer is only an independent check if the reviewer's judgment is genuinely their own.
Deferral of this kind is also exceptionally hard to detect. For work with a correct answer, checking for errors or comparing against a reference answer surfaces the problem. But in judgment calls both options can be defensible, so the work can look sound whichever one was chosen. Asking the experts afterward may not help either, because they never noticed the agent's influence in the first place.
Three directions the authors propose
One is measuring deferral itself. This study could measure it because the researchers randomized the agent's lean, which is rarely possible in real work, and the results suggest people cannot report it accurately about themselves. Two is designing agents differently: elicit the expert's own view before offering one, or make the best case for both sides, though a single analysis can contain so many open decisions that surfacing everything could quickly become overwhelming. Three is changing institutions. Organizations need to decide deliberately which decisions require a person's own informed judgment, and give people the time and resources to form one. Judgment is a skill, and people who stop making these decisions gradually lose the ability to make them.
Full paper: https://www-cdn.anthropic.com/519c17a5d23a794e8b6d50a854048d0b61e0d73a.pdf
AI is taking your work. The more worrying part is that it is taking the part of your judgment you believed was still yours.