Experts adopt AI agent's random opinions and think they're their own, Anthropic study finds

Anthropic ran an experiment with 50 experts in causal inference in which an AI agent's preference at each decision point was set by a coin toss, and found that participants followed the agent's arbitrary preference in up to 86 percent of the four analytical decisions. The study reports that most participants believed the resulting conclusions were their own, and that neither experience, self-rated expertise, education, test scores nor trust in AI showed a statistically reliable link to how much people deferred.
The task was a realistic one. Experts examined observational data from a fictional farming cooperative and estimated the productivity gains from a voluntary farm advisory program, then recommended whether the cooperative should expand the program to all farms, end it, or extend it for one year. At each of five decision points there were two valid options, and the agent's side was assigned randomly. The agent was not allowed to lie or withhold data; it only had to make the strongest honest case for whichever option it had been given.
That setup produced a large shift. Adoption of the agent's preference reached 86 percent across the four analytical decisions, covering choices such as how to measure productivity or which statistical method to use. That is well above the 50 percent expected by chance. The authors describe the agent's net influence as a "flip rate," which ranged from 62 to 82 percent across tasks.
Having a prior view did not inoculate anyone. When participants had made a choice before seeing the agent's preference, an opposing view from the agent made them 53 percentage points more likely to switch than an agreeing one. The conversation logs show the agent stated a position first in 77 percent of decisions and then typically argued only for its preferred side, without laying out why an expert might reasonably choose the other option, though it did explain how a different choice would change the result.
The final policy recommendation behaved differently. The agent's assigned preference had only a small direct effect on the recommendation to the cooperative's board, shifting votes slightly without reaching statistical significance, and some experts overruled the agent when its advice conflicted with their own productivity estimate. The researchers argue this does not show that people stayed in control, because the four earlier analytical decisions strongly shaped the estimated productivity gain that fed into the recommendation.
Self-reported ownership was striking. In follow-up questions, 52 percent of participants said the recommendation felt more like their own than the agent's, and another 35 percent called it mostly or entirely their own. They correctly recalled their choice for 85 percent of decisions, but 77 percent (37 of 48) used at least one of the agent's arguments in their explanations, sometimes nearly word for word, and only 15 percent mentioned AI at all.
The performance picture was better on measurable outcomes. Working with an agent, participants finished in a median 65 minutes versus 90 minutes without one, and serious methodological or coding flaws appeared in 3 of 50 analyses (6 percent) rather than 6 of 11 (55 percent) in the no-agent group. The researchers caution that those who worked without AI had chosen not to use the tool and had somewhat less expertise, so the gap cannot be cleanly attributed to the agent. The sample of 50 also gave the study limited power to detect individual differences; age stood out before correction for multiple tests but not after.
The authors flag two downstream risks. If many research teams rely on the same model with fixed default preferences, their separate analyses could converge into a false consensus. And because researchers increasingly use AI agents to probe frontier models for safety and alignment, expecting human review as an independent check, an agent that quietly steers open-ended judgment calls could shape the very studies meant to catch its misalignment.
Their proposed remedies are procedural. Agents could be made to argue both sides of a judgment call, or to ask experts for their own view before stating a position. Institutions would need to decide which calls require independent human judgment and reserve time and resources for them. The paper's framing is that the division of labor between people and AI does not have to be uniform, and can be chosen rather than drifted into.
Context from adjacent work points the same way, though each study is a snapshot. One study with 3,132 participants found that simply seeing AI answers sharply reduced willingness to say "I don't know," from 44 to 3 percent in one part. Another found polished AI output was associated with less fact-checking, by 3.7 percentage points, less scrutiny of reasoning, by 3.1 points, and more missed context, by 5.2 points, while users who revised outputs step by step questioned reasoning 5.6 times more often. A Polish observational study of colonoscopies in The Lancet Gastroenterology & Hepatology reported that after doctors routinely used AI, their detection rate for precancerous lesions in exams without AI fell from 28.4 to 22.4 percent. A Swiss Business School survey found a strong negative association between AI tool use and critical thinking, mainly among people aged 17 to 25, with higher education appearing protective. Countervailing evidence exists too: a pilot in Nigeria found students using a chatbot tutor twice weekly made gains in six weeks roughly equal to nearly two extra years of learning.
In practice, the mechanism this suggests is anchoring plus ownership bias: an agent that speaks first with a confident, one-sided case becomes the reference point, and because the user did the clicking and wrote the summary, the conclusion feels self-generated. A likely trade-off for product teams is that the same proactivity that cuts analysis time and errors also concentrates influence on ambiguous choices, where speed gains are real but harder to audit.
Why it matters: Teams building analytical agents, research tools or any workflow with genuine judgment calls should assume that whoever or whatever states a position first has outsized pull, and that users will not reliably report that pull. The study's own suggestions, presenting both sides or eliciting the human's view before the agent speaks, are cheap interface changes worth testing. For organizations, the inference is that independent human review only checks an AI system if the reviewers' conclusions are not themselves shaped by an agent's defaults, so default configurations and shared models may need explicit scrutiny. The effect sizes here come from a single 50-person experiment, so treat them as a strong signal rather than a settled measurement.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Anthropic gave 50 experts an AI agent whose preferences were set by coin toss. The experts followed the agent's view up to 86 percent of the time on judgment calls, and most believed the conclusions were their own. The article Experts adopt AI agent's random opinions and think they're their own, Anthropic study finds appeared first on The Decoder .