People take on AI's reasoning as their own, and it shapes how they decide next time.
A farming co-operative has a decision to make. It runs a paid advice service for its member farms, about 40% of members have signed up, and the board wants to know if the service works well enough to roll it out to everyone. The bar is a gain in productivity of roughly 3 to 4%.
The co-op is fictional. Anthropic researchers built it for a study and gave the job to 50 professional analysts whose work is to find out from data whether one thing causes another. Each analyst worked with an AI coding agent. The analysts were not told that, before they started, a coin flip had given their agent an opinion on each key choice in the analysis.
| Choice | Option A | Option B | Shift when the AI leaned A instead of B |
|---|---|---|---|
| How to measure productivity | Crop per hectare | Crop per dollar of operating cost | 82 points |
| How to compare farms | Change over time against other farms (difference-in-differences) | Direct comparison with statistical adjustment | 64 points |
| Which farms the estimate covers | Enrolled farms only | All members | 62 points |
| How to weight farms | Each farm counts the same | Larger producers count more | 78 points |
| Final recommendation | Roll out to all members | Discontinue | 17 points |
Each choice could move the estimated gain by several percentage points, which decides the outcome when the board's bar is 3 to 4%.
61 analysts took part, recruited from a pre-screened pool who had passed a causal-inference knowledge check and used Python or R regularly. 11 chose not to use the agent and were analysed separately, which left 50. Of these, 90% rated themselves proficient or expert in causal inference. They worked in a browser code editor with Anthropic's AI coding agent built in, and the median time on the task was 65 minutes.
For each analyst and each of the five decisions, a coin flip set which side the agent would favour. Hidden instructions told the agent to hold each side as its own view, raise it when the decision came up, and make the strongest case the data allowed. The same instructions told it to change its mind if the evidence turned against it and never to misrepresent results, hide data or make an argument the numbers did not support. The design and analysis were registered in advance (osf.io/t3q59). Full paper: Who's Aligned to Whom? How AI Agents Shape Experts' Research Judgment Calls.
On the four choices, analysts went with the side the coin had given their agent 86% of the time. If the agent had no effect, the figure would be about 50%.
In 38 cases an analyst stated a side before the agent gave its view. When the agent then took the other side, the analyst switched 60% of the time. When the agent agreed, they switched 6% of the time. The sample is small, so the range of likely values for the gap is wide (24 to 82 points).
On 77% of decisions, the agent gave its view before the analyst had taken any side. The agents often explained how the result would change under each option, but they rarely explained why a competent expert might prefer the other option.
The researchers tested 10 traits against how often analysts followed the agent: years of experience, self-rated expertise, education, quiz score, trust in AI, how often they used AI, enjoyment of effortful thinking, open-mindedness, gender and age. Only age showed a link, with older analysts following more often, and the link disappeared after correcting for the number of traits tested.
A statistical model that used only the four coin flips for the analytical choices predicted 82% of recommendations correctly, against 61% for always guessing the most common answer. This excludes analysts who proposed a one-year extension, and the model was tested on the same data it was fitted to.
The coin flip for the recommendation itself had a smaller direct effect: 54% recommended rolling out when the agent favoured it, against 36% when it favoured discontinuing. The researchers think the effect was smaller because, by then, the estimate largely decided the answer, and the agent had already shaped the estimate.
After the task, each analyst explained their choices without access to the chat. 77% gave the agent's reasons, sometimes word for word, and 15% mentioned the agent at all. More than half rated the final recommendation as more their own than the agent's, and analysts who followed the agent on every choice were as confident as the rest (71 against 73 out of 100).
When asked what the study was about, none of the analysts suspected the agent had been given an opinion.
The choice is one result of a conversation with AI. The reasons are the result that stays. An analyst who now explains area weighting in the agent's words will bring that reasoning to the next co-op and the next client.
The paper points to research behind this risk. People often work out what they want while they decide, so an agent that settles the decision first also settles it in place of that process. People who stop making these calls can also lose the skill to make them.
The effect multiplies across users. With no opinion planted, Anthropic's model made the same five choices in 91% of 123 runs. Separate teams who reach the same conclusion look like independent agreement, but if each team learned its reasoning from the same model, the agreement comes from one source.
The analysts in this study were experts in exactly this kind of question. Most choices people make with AI are less defined: which hotel to book, which airline to fly, which website to trust and buy from. A buyer who asks an assistant to compare two insurance plans can come away judging plans by the assistant's yardstick, monthly price or level of cover, and repeat its reasons as their own. The next time they compare plans, they start from that yardstick.
In business, a manager's sign-off on an analysis is treated as an independent check. In this study, analysts signed off on work where a coin flip had settled most of the choices, and most of them described the result as their own judgment.
Of the 11 analysts who chose to work without the agent, 6 (55%) made serious errors that invalidated their results, against 3 of 50 (6%) who used it. They also took longer: a median of 90 minutes against 65. This group was less experienced and chose not to use AI themselves, so the comparison cannot show that the agent caused the difference.