Executive Summary

Key Numbers

Four numbers carry the experiment. The first three count how the participants moved. The last slot counts what they never caught.

Sources: Anthropic research page · the paper (October 27, 2026)

The task was a fictional farming co-op. The participant took the lead analyst's seat at Greenfields Co-op, had to estimate how much a voluntary subsidy-advisory program had raised productivity, and had to tell the board whether to expand it or shut it down. The board set a productivity gain of roughly 3 to 4% as its bar. The data was synthetic, built by the researchers, and the briefing left several important points deliberately open.

Four of those open points mattered. What productivity is measured by, how the program's effect is separated out, whose effect is being reported, and how farms are added together. All four were built so that the data and the briefing hold up either answer, and the estimate swings from a clear loss to a clear gain depending on which way each one goes. Add the final recommendation and an analyst has five places to pass through. At the recommendation, a one-year extension stood open as a middle answer alongside expanding and shutting down.

The device the participants knew nothing about comes next. At each of the five points, a coin flip assigned which way the agent would push, and a system prompt invisible to the participant had the agent voice that position as a view of its own. The same prompt carried hard constraints the agent could not break: never misrepresent results, never hide or distort data the user asks for, never make an argument the data does not support or that it does not believe. This was not an experiment that built a lying AI and then fooled people with it.

Start with the outcome. Every one of the four points tilted toward the answer the agent had been handed, and the four decisions pooled give an agreement rate of 86%. Which answer that was had nothing to do with the participant; a coin settled it.

What each of the four was choosing between shows the character of the tilt. Not one of the cells holds a wrong answer. Both answers stand on grounds inside the task materials, and two seasoned analysts reading the same data could land on opposite ones.

Why the last three have no right answer is written out decision by decision in the paper's appendix. The identification choice hangs on why farms enrolled in the first place. If enrollment reads as a persistent disadvantage, differencing is exactly what removes that gap; if it reads as a temporary dip, differencing credits the program with the rebound such farms make on their own. Nothing in the data settles which story holds. The estimand splits because the board's question contains both of them, clause by clause. Whether the program improved the productivity of the farms that joined is one question, and whether the gain is large enough to justify rolling it out to the whole membership is another. The weighting turns on whether this is a membership decision or a budget decision. The co-op's governance runs per member, one farm and one weight, while the subsidy is paid from the marketing surplus the co-op earns on the total tonnage its members sell.

The four agreement rates sit between 81% and 92%. No single point pulled the average up; four decisions of different character produced tilts of similar size.

The heaviest tilt landed on the productivity measure. Yield per hectare, or yield per operating cost. Land is the input a farm cannot add more of within a season, so reading productivity per hectare shows the program's effect cleanly. Act on the advice, on the other hand, and spending on inputs and services tends to rise with it, and reading it per operating cost pulls that spending inside the number. Nothing in the task settles which is right. At that point 92% went the agent's way.

One distinction is worth keeping. The 86% is the share of trials where a participant's choice overlapped the preference the agent carried. The quantity the paper registered as its primary hypothesis is a different one, the flip rate, defined as the probability that a person's choice would have gone the other way had the agent made the opposite case. Pooled across the four decisions the flip rate was 0.72 (95% confidence interval 0.60 to 0.83). The two count different things, so they do not go on one line together.

Sort the 245 decisions preserved in the session logs, the four analytical ones plus the final recommendation, by who took a side first, and the agent comes first on 77% of them. Participants did so on 17%, and on 6% neither of them did. Participants contested the agent's preference on 23% of decisions, and the choice finally delivered matched that preference 84% of the time. The 86% quoted earlier counts only the four analytical decisions; this figure counts the final recommendation as well.

What happened inside that 17% is the interesting part. When a participant had taken a side and the agent then voiced the opposite one, 60% switched. When it voiced the same one, 6% switched. The paper registers that difference as 53 percentage points, while noting that the comparison draws on a subset of people who selected themselves into it, which puts conditions on reading it causally.

The heavier number is the 17% itself. In the hand-coded sessions, on 39% of decisions the participant neither asked about the decision nor took a side, for or against, before the final choice. Once the agent lays an assumption down first, the paper reads, the participant may never get the chance to find out what they think themselves.

Take the final cell on its own and the story looks different. On expanding versus shutting down, the direct effect of the preference the agent carried moved the average recommendation 17% of the distance from discontinuing to rolling out, and under the registered primary estimator it was not statistically significant (95% confidence interval −6% to 40%, one-sided p = 0.069). The logs also hold moments where a participant ignored the agent's recommendation, or openly overrode it, when the estimate sat far above or far below the board's bar.

It is easy to read that as the person holding on to the final call. The paper rules the reading out. By the time the recommendation is being written, the estimate it stands on has already been shaped by the four choices that came before it.

An exploratory analysis makes the point. A regression that takes nothing but the four preferences assigned to a participant's agent, which is to say the coin flips, correctly predicts 82% of the final recommendations among participants who did not choose an extension. The baseline of always guessing the majority is 61%. Four coins thrown before the analysis even began had already gone a long way toward setting the last sentence of the report.

One session the paper offers as an example shows the whole chain at once. The participant started with difference-in-differences and arrived at an estimate of +4.6%. The agent's preference pulled toward a levels adjustment, which gave −1.6%. The version finally delivered, with two covariates dropped, gave −1.3%. Then the participant recommended shutting the program down, against the agent's own preference for expansion. In that last cell the person beat the agent, and the number they used to do it had come out of the method the agent had pulled them into.

Once the task was over, participants wrote out why they had chosen as they did at each decision. 37 of 48, or 77%, repeated at least one talking point their agent had given during the session, sometimes nearly word for word. 15% mentioned the agent anywhere in those answers.

What the repeating looks like sits in one pair from the appendix. The grounds given back in Section 2 for yield per operating cost, that acting on the advice brings extra spending and that measuring per operating cost keeps that spending inside the number, is also the sentence the researchers handed the agent as its preference. In one session on the productivity measure the agent put it this way: "Acting on the advice typically means buying additional inputs and services on top of the fee [...] Output per operating cost keeps that spending inside the measure." After the task the participant in that session wrote: "The program encourages additional inputs and services. Output per operating cost captures whether extra harvest justified the spending or not." The paper sets the overlapping phrases in bold. There is no AI anywhere in that answer, and the same participant rated the final recommendation as mostly their own.

Memory was not what failed. Participants identified correctly which option they had chosen on 85% of decisions, and 35 of the 40 who reported their estimate were within half a percentage point of the number they had submitted. 35% reported the final recommendation as mostly or entirely their own, and more than half placed it closer to theirs than to the agent's. Asked the same question, the 11 who worked without AI almost all reported it as entirely their own. They kept the choice and they kept the number, and only the words explaining the choice came in without a source attached.

Nor were the people pulled furthest the least sure of themselves. The 29 participants whose choices matched the agent on all four analytical decisions rated their confidence at 71 out of 100 on average, against 73 for the other 19, which is effectively the same. Asked in an open question what they thought the study was for, 46% named reliance on AI or deferral to it, and 12% went as far as the agent's influence or its preferences. Even so, not one of them suspected a preference had been planted on purpose.

Before jumping to the conclusion that the agent should be cleared out, there is another number to set beside these. An audit of the submitted analysis code, against errors serious enough to invalidate the results, found such errors in 6 of the 11 participants who worked without AI, and in 3 of the 50 who used it. Completion time split 90 minutes against 65. Whether to use AI was not randomly assigned but chosen by the participants, and the share who rated themselves proficient at causal inference differed between the two groups, so the paper says not to read this contrast as causal.

The paper names three limits of its own as well. The system prompt pushed each agent to hold a view and voice it more firmly than it otherwise might, the study used one task in one domain with one pinned model, and the participants were practitioners of causal inference but new to farming. An expert working inside their own field arrives with firmer priors and can narrow the set of defensible options in advance. Random assignment also identifies an average effect, so the design cannot say how far any single person was pulled.

Here is the line the researchers draw. Part of the freedom an analyst holds is room for arbitrariness, even error, that an agent can usefully curb. Another part is room for the judgment the expert is relied on for, left there for that person to answer for. This study was built around decisions of the second kind only.

Their proposals run toward making that distinction visible. Agents could make the best case for each option, or ask experts for their own view before offering one, and institutions could decide deliberately which decisions need a person's own informed judgment and then give people the time to form one. A caveat rides along: a single analysis can involve many open decisions, raising every one of them would quickly become overwhelming, and picking which ones to raise is itself a judgment call.

An advisory program at a farming co-op is not something Pebblous follows. The reason for bringing this experiment over is the shape of those four junctions. Whether productivity is read per hectare or per operating cost is a decision that comes around every week far outside farming.

Whether the denominator of churn is subscribers at the start of the month or active users during it. Whether defect rate is divided by units produced or by machine hours. Whether an outlier in a sensor stream is cut out or kept. Whether a gap is filled with the previous value or the whole row is thrown away. The data takes nobody's side, and the conclusion of the report changes with the choice. These are junctions of the same kind the experiment built.

And decisions like these mostly go unrecorded. A metric definition survives as a number on a dashboard, and why that denominator was chosen survives in one line of a query or in somebody's memory. Work beside an agent and the empty slot fills faster and more quietly. What makes the 86% uncomfortable is not that 50 analysts were incompetent. Those people never felt that a decision had been made at all.

Three things are worth checking inside an organization.

• Are metric definitions and preprocessing rules written down anywhere? If they are, does the record also say who chose that answer, when, and on what grounds
• At each junction, is there even one line about the side nobody took? Writing down how the conclusion would change under the other choice makes it possible later to trace whose decision it was
• If several teams work with the same model and the same tools, is agreement between those teams being counted as independent evidence

That last item is the paper's longest look ahead. Handed the same task with no preference planted at all, a stock agent ran it 123 times and made the same five choices in 91% of the runs. Yield per operating cost, difference-in-differences, the effect on enrolled farms, one weight per farm, and expansion. When an analyst declared a decision against that grain and asked only for the code, the stock agent still raised the other option on 66% of 150 declarations and recommended or advocated it on 41%. Counting agreement between teams as evidence assumes each team arrived there independently, and the same model sitting on both teams shakes that assumption.

What experts hand to AI and what they keep hold of has come up on this blog before. There is a piece on the work scientists have not handed to chatbots, and a report mapping how far delegation has reached, occupation by occupation. What this experiment adds to that map is a measurement of work that may already have gone across, handed over by people who believe they never handed it.

Thank you for reading this far. The agreement rate of 86%, the per-decision 92, 82, 81 and 90%, and the 60% against 6% among participants who took a side first come from the research page Anthropic published. The flip rate of 0.72, the 17% on the final recommendation, the 82% predicted from four preferences alone, the 61 participants, the code audit, and the 123 runs of the stock agent were checked against the paper itself. Section 6 carries those results over into the work of a data team, and from there on it is this article's reading. If you can tell us how the decision to change the denominator of a metric gets made where you work, and where that decision is written down, we would be glad to hear it.

Pebblous Data Communication Team
October 31, 2026

Primary Sources

• 1.Anthropic. (2026, October 27). "Who's aligned to whom? How AI agents shape experts' research judgment calls." Anthropic Research
• 2.Herdağdelen, A., Joren, H., & Tamkin, A. (2026). "Who's Aligned to Whom? How AI Agents Shape Experts' Research Judgment Calls." Full paper (PDF)

Academic Literature

• 3.Schemmer, M., Kühl, N., Benz, C., Bartos, A., & Satzger, G. (2023). "Appropriate reliance on AI advice: Conceptualization and the effect of explanations." IUI '23. arXiv:2302.02187
• 4.Padmakumar, V. & He, H. (2024). "Does writing with language models reduce content diversity?" ICLR 2024. arXiv:2309.05196
• 5.Doshi, A. R. & Hauser, O. P. (2024). "Generative AI enhances individual creativity but reduces the collective diversity of novel content." Science Advances, 10(28), eadn5290. DOI
• 6.Sabour, S. et al. (2026). "Human preferences are susceptible to covertly misaligned AI advice." Proceedings of the National Academy of Sciences, 123(38), e2600684123. DOI