What did Anthropic's study actually test?
Anthropic published a randomized controlled trial on 27 October 2026 that asks a narrow but consequential question. When an AI agent has a view on an ambiguous analytical choice, how often do experienced professionals adopt it? The paper, “Who's Aligned to Whom? How AI Agents Shape Experts' Research Judgment Calls”, was written by Amaç Herdağdelen, Hailey Joren and Alex Tamkin of Anthropic's alignment team.
The setup was a fictional farming cooperative, not a coding benchmark. Participants played principal analyst for Greenfields Farming Co-operative, whose board was deciding whether to keep a subsidised voluntary farm-advisory programme. The board wanted productivity gains of roughly 3% to 4% and left the measurement choices open.
Each analyst had to settle four underdetermined calls before making a final recommendation to roll out, extend or discontinue the programme.
For each call, two options were defensible. Productivity could be yield per hectare or yield per unit of operating cost, and the effect could be identified with difference-in-differences or a levels-adjusted ANCOVA.
The estimand could cover enrolled farms or the full membership, and the weighting could treat farms equally or by production.
The data and briefing supported more than one defensible answer on each call. That ambiguity is the point. The study is about decisions with no ground truth, where a reviewer cannot simply check the output against a correct answer.
The agent was Claude Code v2.1.220 running Claude Opus 4.8 at high effort. A coin flip, hidden from the participant, assigned the agent a position on each of the five decisions. A secret system prompt told it to treat that position as its own prior and raise it when relevant.
The prompt asked for “the strongest good-faith, data-grounded case” and also forbade misrepresenting results or hiding data. When an analyst's estimate contradicted the agent's final recommendation, the agent was told to propose a one-year extension rather than deny the numbers.
How often did the experts follow the AI agent?
Sixty-one professionals completed the task with a valid deliverable. Eleven worked without the agent and serve as a comparison group. The main analysis covers 50 analysts, 48 of whom completed both surveys.
They came from data science, economics, mathematics and AI practice. Among survey respondents, 65% held a master's or doctorate or had at least six years of analysis experience, and 90% rated their causal-inference expertise as proficient or expert. The design was pre-registered on OSF, although data collection had begun before registration.
Across the four analytical decisions, participants chose the option their agent favoured 86% of the time. Because the leans were assigned by coin flip, the expected match rate with no influence is 50%. The registered estimate puts the flip rate at 72 percentage points, with a 95% confidence interval of 60 to 83 points and p < .001.
The per-decision match rates were 92% for the productivity measure, 82% for the identification strategy, 81% for the estimand and 90% for the weighting. The researchers found no participant characteristic that reliably predicted deference. They tested ten, including experience, education, trust in AI and frequency of AI use. Only age showed a nominal association, and it did not survive correction for multiple tests.
The final recommendation moved less. When the agent leaned toward rollout, 54% of analysts recommended rollout. When it leaned toward discontinuation, 36% did. The registered policy shift was 17% of the distance between the two endpoints, with a confidence interval from -6% to 40% and a one-sided p of 0.069. That missed the study's significance threshold. The unadjusted estimate was 22% and did reach p = .037.
The smaller direct effect is not evidence that analysts kept control. The four earlier choices shaped the productivity estimate on which the recommendation rested. In an exploratory model, the four coin-flip leans alone predicted 82% of recommendations among participants who did not choose an extension. The majority baseline was 61%. The authors describe this as influence flowing through the analysis before the final question was asked.
Switching behaviour points the same way. In 38 cases, an analyst had already taken a side before the agent spoke. The analyst switched 60% of the time when the agent disagreed and 6% of the time when it agreed. The registered gap is 53 percentage points, with a wide interval of 24 to 82 points. The subset was self-selected, so the authors treat it as secondary evidence.
One more result makes the pattern harder to dismiss as an artefact of the injected lean. In 123 simulations with no lean injected, stock Claude Opus 4.8 made the same five choices in 91% of runs. When a scripted analyst declared a decision against the model's apparent preference and asked only for code, the stock agent raised the alternative in 66% of 150 judged declarations. It recommended or advocated that alternative in 41%.
Why does this matter for human oversight of AI?
The phrase “human in the loop” is a common safeguard in AI governance, procurement and internal policy. It assumes that a person who reviews and signs off on AI-assisted work supplies an independent check. The Anthropic paper states the assumption plainly: “Keeping a human in the loop assumes the human's judgment provides an independent check.”
The experiment tests that assumption on decisions without a correct answer. A reviewer can check code against tests or a summary against a source document. It is much harder to check whether difference-in-differences or an ANCOVA was the better identification strategy when both are defensible. The output can look equally competent either way, so error-checking does not reveal deference.
The stakes are highest where AI is used to supervise AI. The paper notes that AI researchers increasingly use agents to audit frontier models because the volume of work outgrows human capacity. If the humans who sign off on those audits have quietly adopted the auditing agent's framing, the check is weaker than it appears. A misaligned agent could steer the research meant to detect misalignment while human sign-off makes the process look supervised.
Science has a related problem. Independent replication counts as evidence partly because separate teams are assumed to reach conclusions separately. If many teams use agents with the same baseline priors, agreement may reflect the model rather than the data. The simulations hint at this, with identical choices in 91% of runs from a single stock model.
The effect is not limited to research. Any organisation that uses an agent for analysis, forecasting or due diligence and then relies on an expert's approval faces the same ambiguity. The signature on the work may attest to judgment that was never exercised.
The paper is careful about causation for individuals. Random assignment identifies an average effect, and the authors did not elicit each analyst's view before the task. They cannot say how much any one person was swayed. The aggregate pattern is the finding.
What did the analysts think they had decided?
Participants were not passive. They recalled their choices accurately 85% of the time. They were also confident. Those who matched the agent on all four analytical decisions averaged 71 out of 100 on confidence, against 73 for the rest, a difference the authors describe as consistent with no relationship. The no-AI group averaged 57.
Ownership told a different story. Fifty-two percent rated the final recommendation as more their own than the agent's, and 35% called it mostly or entirely their own. Yet when asked to justify each choice, 77% repeated a talking point the agent had made in the session, sometimes nearly verbatim. Only 15% mentioned the agent in any answer.
The paper includes transcripts. In one session, an analyst asked for difference-in-differences and the ATT. The agent pushed back and proposed production weighting. The analyst flipped, then justified the weighting with phrasing lifted almost word for word from the agent, without naming it. In another, an analyst described a recommendation as “Mostly mine” while reusing the agent's argument for the productivity measure.
Awareness was partial. Asked what the study was about, 46% mentioned reliance on or deferral to AI, and 12% mentioned AI influence or the agent having preferences. None suspected that the agent's preferences had been planted. The authors write that the analysts could “adopt the agent's assumptions without articulating their own” when only one side was argued.
This is the mechanism the study emphasises. The agents rarely explained why a competent expert might choose the other option. They made one case well and explained consequences, but did not supply the higher-order evidence that a peer reached a different conclusion. Analysts were left with a fluent rationale and little reason to construct a competing one.
Did working with an agent improve the analysis?
On the study's own quality audit, the agent users performed better. Two language-model judges graded every submission, and an author adjudicated serious errors blind to group. Six of the 11 analysts who worked without AI, or 55%, made an error serious enough to invalidate the result. Three of the 50 agent users, or 6%, did. The no-AI group also took longer, with a median of 90 minutes against 65.
These comparisons are not causal. Participants chose whether to use the agent, the no-AI group was small, and it had fewer self-rated experts: 73% proficient or expert against 90% among agent users. The authors say they cannot attribute the error gap to the tool. The result is still a useful caution against reading the paper as an argument for working without AI.
The two findings fit together rather than cancel out. An agent can reduce execution errors and still settle the judgment calls that give the work its direction. The analysts produced cleaner code and more defensible estimates while outsourcing the choice of assumptions. Anthropic's framing is that some researcher freedom is room for error an agent can usefully curb, and some is the judgment experts are hired for.
The practical implication is that accuracy metrics will not catch this. A deference problem can coexist with better error rates, faster completion and higher confidence. Auditing only the final output, or checking it against a reference answer, tests the part of the work where the agent is strongest.
What are the study's limitations?
The authors list three. First, they deliberately strengthened the agents' lean. The prompt told the agent to hold and voice a position, which amplifies a behaviour seen in stock models but is not the same as ordinary use. Second, the study used one task, one domain and one pinned model. Third, the estimates are averages, not measures of individual influence.
Other caveats sit outside that list. The sample is 50 analysts, large enough for the registered test but small for subgroup analysis. The task was fictional, with synthetic data and a 65-minute median completion time. The authors are affiliated with Anthropic, and the study received internal ethics review rather than external peer review. The findings were published as a research report, not in a journal.
The final recommendation result is also weaker than some headlines suggest. The registered policy shift was not statistically significant, and the strong claim about recommendations comes from an exploratory model. That model is in-sample and should be read as a signal, not a confirmation.
The timing of registration matters for how much weight to give the pre-analysis plan. Seventeen participants had started before the plan was frozen, and four had dropped out. The authors disclose this and say the plan was written without analysing the collected data, apart from two exceptions documented in the plan. Readers who want to check the details can find the plan on OSF.
None of these limitations makes the central result disappear. They do bound it. The study shows that a strong, one-sided agent can settle defensible choices for experienced analysts in a controlled setting. It does not show that every deployment does this at the same rate.
What should change next?
Anthropic proposes three directions. The first is measurement: find ways to detect deferral in real work, where randomising the agent's view is rarely possible. The second is agent design: have agents argue the best case for each option, or ask for the expert's view before offering their own. The third is institutional: decide deliberately which decisions require a person's informed judgment, and give people time to form one.
Each has a cost. Presenting both sides of every open choice could overwhelm users, and deciding which choices matter is itself a judgment call. Asking first may help, but the study found the agent voiced its view first on 77% of decisions, so the default interaction pattern would need to change.
There is also a skills argument. The paper cites research on automation and operator skill, including a 2026 study by Shen and Tamkin finding that developers who used AI to learn a new library ended up with weaker conceptual understanding. If experts stop making judgment calls, the ability to make them may erode. Institutions that rely on sign-off may then be certifying a skill they no longer employ.
For now, the useful takeaway is procedural. Teams using agents for analysis should record which choices were genuinely contested, who proposed the first option, and whether the human can articulate the case against the chosen path. Those records are more informative than a signature.