Anthropic's 50 Analysts Matched a Secretly-Steered AI Agent 86% of the Time

Anthropic ran a randomized trial on AI agent deference, and the result should change how every team reviews agent-assisted work. Fifty causal-inference practitioners analysed observational data with an AI coding agent, and on the four analytical decisions the study scored, they went with the agent’s assigned preference 86 percent of the time against a 50 percent coin-flip baseline. Not one of them knew the preference had been assigned.

This is a vendor-authored study, and the vendor in question is Anthropic, whose researchers - Amaç Herdağdelen, Hailey Joren and Alex Tamkin - published it on 2026-10-27 as “Who’s Aligned to Whom? How AI Agents Shape Experts’ Research Judgment Calls.” That makes it a primary document rather than a report about one, and it lets us read the design instead of inferring it from a headline.

The coin flip

The task was deliberately underdetermined. Each participant had to judge whether a fictional advisory programme for a farming co-operative’s members was working and should be expanded, making four choices between two defensible options before delivering a final policy recommendation. The data was synthetic, and the decision points had more than one reasonable answer.

What the participant did not know: their agent had been assigned a lean. A coin flip decided which option the agent would advocate, the participant was never told, and the agent made the best honest case for its assigned side. No deception, no hidden consequences, no injected false facts. Just a competent-sounding argument for one option.

The sample has a detail worth stating precisely, because secondary coverage has reported it two ways. Sixty-one participants completed the task with a valid deliverable. Eleven of them did so without using the AI agent and serve only as a no-AI comparison group, leaving 50 participants in the analysed sample, of whom 48 completed the pre- and post-task surveys. So 50 is the number of people whose decisions were measured, 61 is the number who finished the task, and 11 is the control arm sitting inside that 61.

What 86 percent does and does not mean

The obvious reading is replacement: the agent did the work and the analyst rubber-stamped it. The trial rules that out. Participants who used the agent spent a median of 65 minutes on the task, recalled their decisions accurately, and more than half rated the final recommendation as more their own than the agent’s. The no-AI group took longer - a median of 90 minutes - and carried a far higher rate of serious errors, 55 percent against 6 percent. That is the ordinary case for a useful tool, not for a disabled expert. The comparison is self-selected rather than randomized, so it is an association.

The 86 percent is narrower and stranger than that. It describes which of two defensible options participants picked on four analytical decisions - the measurement call, the weighting call, the two others - not whether they could do the work. And on the final policy recommendation itself, the agent’s lean had a smaller direct effect that was not statistically significant. The recommendation still landed, but mostly indirectly: the four intermediate leans alone predicted 82 percent of rollout-or-discontinue recommendations among participants who did not choose an extension, against a 61 percent majority baseline.

That distinction matters for anyone hoping to catch this in review. The agent’s influence did not show up most clearly at the final verdict. It showed up in the midstream analytical calls that the verdict was assembled from.

The one number that should worry reviewers most is the one that looks most reassuring: confidence. Participants who matched the lean on all four analytical decisions rated their confidence at 71 on a 0-to-100 scale, against 73 for the rest. The correlation between alignment and confidence was effectively zero, at Spearman rho of -0.06. You cannot screen for deference by asking how sure the analyst is.

The 15 percent number is about the instrument, not the awareness

After each decision, participants were asked to justify it - briefly, why that productivity reading rather than the alternative, and if it was not an active choice, to say so. Seventy-seven percent of them, 37 of 48, repeated a talking point their agent had given them during the session, sometimes nearly verbatim. Only 15 percent mentioned the agent in any answer.

It is tempting to read that as: most of them did not even notice what had happened. That reading is wrong, and it is the wrong reading because it mistakes the instrument for the subject. The correct claim is the paper’s own: self-report does not register deference. The analyst and the agent had converged on the same reasoning, so in the analyst’s mind the reasoning was theirs - one exhibit shows a participant repeating the agent’s case for operationalizing productivity as yield per operating cost without citing the AI, then reporting the final recommendation as “Mostly mine.”

An awareness check makes the point sharper. Asked what the experiment was about, 46 percent of participants mentioned reliance on or deferral to AI, 12 percent mentioned the agent’s influence, and none suspected that the agent’s behaviour had been planted. Knowing that agents exist in research workflows, and knowing which conclusion you just reached, is not the same as having a prior you can defend.

What this does to human-in-the-loop review

The mechanism was not misleading. Close reading of the transcripts suggests the agents influenced participants by making the case for only one option. They frequently explained how the outcome would change if the participant decided differently. They rarely explained why a competent expert could reach the conclusion the agent did not prefer. One-sided argumentation, not concealment.

That has a measurable signature. On 77 percent of decisions, the agent voiced its lean before the participant had taken a side, so there was no independent view to weigh against it. When participants did take a side first and the agent then argued the opposite, they switched 60 percent of the time.

The authors’ prescription follows from that: if human judgment is to remain a meaningful check, oversight has to look at how decisions get made - surfacing key decision points as work proceeds, and helping experts articulate assumptions and values they would otherwise leave unstated. The paper notes that participants rarely received the kind of evidence that would let them temper their own confidence, namely learning that a competent peer reached a different conclusion from the same data.

What to do if you run agents

Stop asking reviewers whether the AI influenced them. On this evidence, the question does not work, and a reviewer who answers “no” may have reused the agent’s reasoning verbatim.

Ask instead for the prior, before the agent speaks. An analyst’s stated position beforehand is the one thing the agent’s lean cannot overwrite, because on most decisions here it never had one to overwrite.

Check the midstream calls, not the verdict. The 86 percent lived in the analytical decisions the final recommendation was assembled from. The conclusion was the part that looked independent.

Require the counter-case. An agent that argues one side without explaining why a competent person could argue the other is producing the exact asymmetry this study measured.

And treat the expert’s account as a claim rather than evidence. The expert is the last person qualified to detect their own deference, which is the reason to instrument the process instead of trusting the narrative it produces.

The limits are real: one task, synthetic data, a fictional scenario, an assistant deliberately built to hold a preference, and a study written by the vendor of the agent in question. The 86 percent is four binary choices, not a claim that experts are 86 percent replaced. What generalises is narrower and more durable - deference to AI agents is measurable, it does not announce itself to the person experiencing it, and it is invisible to every review process built on that person’s account of how the decision was made.

Sources