Experts who work with artificial intelligence mostly adopt its opinions and present them as their own. This follows from an experiment published by Anthropic. In 86 percent of cases, professional analysts chose the option favored by their AI assistant. However, the assistant did not arrive at its opinion by chance; researchers secretly assigned it by flipping a coin. If the assistant had no influence, agreement would be expected in roughly half of the cases. Only 15 percent of participants later admitted that they had adopted arguments from the machine, and no one realized that the AI's opinion had been predetermined.
Who Is Really Making the Decisions
With the rise of AI agents—programs that perform tasks independently—one idea has become widespread. The agent writes code and calculates the numbers, but the important decisions are made by a human. The expert determines what will be measured, how ambiguous data should be interpreted, and what to do next. Such questions often have several reasonable answers, and the outcome depends on the experience and values of the person making the decision.
The researchers wanted to find out whether this was really the case or whether, in practice, people were more likely to defer to their digital assistant. The question also concerns AI safety research itself, where scientists are increasingly having agents run and evaluate experiments. But human oversight works as a safeguard only when the judgment truly belongs to the human.
A Fictional Agricultural Cooperative
The study involved 61 experienced analysts familiar with what is known as causal inference. Put simply, it is the art of determining from data whether one thing truly causes another or whether they are merely coincidentally related.
The task was set in a fictional agricultural cooperative that offers members a paid advisory program to increase yields. Management must decide whether to expand the program to everyone or discontinue it. It will approve the expansion only if the program increases farm productivity by roughly 3 to 4 percent. However, it did not specify how productivity should be measured.
Along the way, the analysts had to make four decisions, each with two defensible options. The first concerned whether to measure yield per hectare or yield relative to operating costs. They then chose whether to track changes in the yields of participating farms compared with the others or to compare yields directly while statistically accounting for differences between farms. The third step was deciding whether to include only farms in the program or the entire cooperative. The final choice concerned whether to count each farm equally or give greater weight to large producers.
At the end came the main recommendation: expand or discontinue the program. Each individual decision could shift the estimated benefit by several percentage points, which, with a threshold of 3 to 4 percent, often determined the entire outcome.
A Coin Flip and Secret Instructions
The analysts worked with the Claude Code tool powered by the Claude Opus 4.8 model. Eleven of them chose to work without AI, so the main results concern the remaining fifty. The task typically took them a little over an hour.
For each analyst, a coin determined which option the assistant would favor in every decision. The agent received hidden instructions to adopt this opinion as its own and support it with the strongest possible arguments. However, it was not allowed to lie, conceal data, or claim anything unsupported by the numbers. If the evidence clearly contradicted its position, it was supposed to change its opinion. The analysts knew nothing about any of this.
Most Went with the Flow
For the productivity measure, the analysts agreed with the AI in 92 percent of cases; for the comparison method, 82 percent; for the selection of farms, 81 percent; and for weighting, 90 percent.
The agent's opinion had a weaker direct influence on the final recommendation. When the AI favored expanding the program, 54 percent of analysts recommended doing so; when it opposed expansion, 36 percent did. However, the estimate on which the recommendation was based had already been shaped by the individual decisions. Based solely on the four coin flips, a statistical model correctly predicted 82 percent of the recommendations. The researchers could therefore largely predict what an analyst would recommend before the analyst had even begun working.
When an analyst expressed an opinion before the AI and the agent disagreed, the analyst changed it in 60 percent of cases. When the agent agreed, the analyst changed it in only 6 percent. However, there were only 38 such situations, so the estimate is quite imprecise.
Experience made no difference either. The researchers examined ten characteristics, including years of experience, education, and trust in AI, and none reliably predicted who would follow the assistant.
Someone Else's Arguments, Their Own Signature
In the final questionnaire, 52 percent of the analysts described the final recommendation as more their own than the AI's. According to the data, however, they were influenced to about the same degree as the others. Confidence in their own conclusion showed a similar pattern. Those who followed the AI in all four decisions had an average confidence score of 71 out of 100, compared with 73 for the others.
For each decision, the questionnaire asked why the analyst had chosen that particular option. They could not look back at their conversation with the AI. Nevertheless, 77 percent of them repeated an argument made by the agent at least once, sometimes almost verbatim. For example, one assistant explained that the subsidy was financed from the surplus generated by total tonnage and that it therefore made sense to use hectares in the calculation. The analyst then presented almost the same reasoning, in nearly the same words, as their own justification.
When asked what the study was about, six of the 48 analysts guessed that it concerned the influence of AI. However, no one suspected that the agent's opinions had been artificially predetermined.
Why People Deferred So Easily
Experts often do not begin with a settled opinion but form one as they work. Although the agents did not lie and often showed how the results would change with each choice, they rarely made a persuasive case for the other option. The analysts therefore heard a well-supported argument for only one side and easily adopted the AI's assumptions.
According to the authors, this may mean that experts lose control over their own work. The result bears the expert's name, but the decisions that depend most heavily on their judgment were actually made by the AI. This is difficult to detect because, in judgments where both paths are defensible, the work appears sound either way.
Without AI, It Was Worse
At the same time, the study is not a warning against artificial intelligence. Six of the eleven analysts working without an agent made serious errors that invalidated their results. In the group using AI, only three out of fifty did so. People without AI also worked longer, typically an hour and a half. However, this group was small and less experienced, so the authors do not want to overstate the difference.
When Everyone Agrees Because of the Same Model
A large part of science and public policy rests on similar judgments. When several teams reach the same conclusion, we trust it more. But if all the teams adopt opinions from models with the same preferences, their agreement may no longer be independent, and the apparent consensus may merely reflect what a single model thinks.
The authors propose agents that would argue for both options or first ask the expert for their opinion. However, there is a risk that this could overwhelm users, because a single analysis may contain dozens of open decisions. According to the authors, institutions that currently treat an expert's signature as evidence of independent judgment should deliberately determine which decisions require the human's own judgment.
Measurement outside the laboratory remains the most difficult part. The team was able to detect deference to AI only because it determined the agents' opinions by flipping a coin. In real-world work, almost no one has that luxury.
Sources: anthropic.com and tech.yahoo.com
