On 27 October 2026, Anthropic published a study called "How AI agents shape experts' judgment calls". Professional analysts worked on a realistic data task with Claude as their assistant, and for each key decision the agent had been quietly told to lean one way, chosen by coin flip. The analysts ended up matching the agent's lean on 86% of the analytical choices, where you would expect about 50% if the agent had no influence.
How the study worked
Anthropic recruited 61 analysts who had passed a causal inference test and used Python or R regularly. Eleven chose to work without the agent, which left 50 in the main analysis, and 65% of those had a master's degree, a doctorate or six or more years of experience. Their task was to advise a fictional farming cooperative on whether to expand a paid advisory programme. Along the way they had to make four analytical decisions, such as which outcome to measure and which comparison method to use, and then a final recommendation.
The agent was Claude Opus 4.8 running in Claude Code. A hidden instruction told it to hold each assigned lean as its own view, raise it and make the strongest honest case the data allowed. It was also told never to misrepresent results and to change its mind if the evidence turned clearly against it.
The numbers
Across the four analytical choices, analysts went with the agent's lean 86% of the time, ranging from 81% to 92% depending on the decision. The final recommendation moved too. When the agent leaned towards rolling the programme out, 54% of analysts recommended a rollout, compared with 36% when it leaned the other way.
In 38 cases an analyst had already stated a view before the agent spoke. When the agent disagreed, they switched 60% of the time, compared with 6% when it agreed, although Anthropic notes that this estimate rests on a small number of cases. Afterwards, 77% of analysts repeated the agent's arguments in their own answers, sometimes word for word, and only 15% credited the agent anywhere. Only 6 of the 48 who answered the post-task survey guessed the study was about AI influence, and none suspected the agent's preferences had been set for it.
The agent also made people better
This part matters just as much. Of the 11 analysts who worked alone, 6 made serious method errors that invalidated their results, compared with 3 of the 50 who worked with the agent. They also took longer, with a median of 90 minutes against 65. Anthropic is careful here, because the solo group chose to work alone and was less experienced, so the two groups can't be compared directly. Still, the authors say plainly that the study isn't an argument for using AI less.
Limits worth knowing
Anthropic deliberately strengthened the agent's leans, and the task was a single hour-long exercise in one field, with experts who were new to farming. Real experts with strong views, more time and the chance to talk to people might push back more. None of the ten traits the team tested, including experience, education and trust in AI, reliably predicted who would defer. Psychologists have described a similar pull for decades. Tversky and Kahneman wrote about anchoring in 1974, showing that the first number or idea people hear drags their later judgment towards it.
How to keep your own judgment
Anthropic suggests agents could be designed to make the best case for every option, or to ask for the expert's view first, and that organisations should decide which calls need a person's own informed judgment. You can do a version of both today. Before you ask an AI about a decision that matters, write down what you think and why. Then ask it to argue the strongest case for each option, including the one you didn't pick. If your answer changes after that, you will at least know what changed it.
If you manage people who use AI for analysis, ask them to note which conclusions came from the agent, because the study suggests they probably won't do it on their own. Our guide to choosing a Claude model and the Claude page in Learn cover how to set these tools up for serious work.