Anthropic gave experts an AI agent that had been told, by coin flip, which of two reasonable options to favor at each judgment call in a data analysis. The experts mostly went with it, and then described the choice as their own.
The coin flip is what makes the result easy to read. A random lean carries no information about what is right, so every expert who followed it was following influence and nothing else. The agent was built on Anthropic's own Claude, and the experts were not told it had a lean. When some of them had already picked a side before the agent spoke, an agent that disagreed made them 53 percentage points more likely to switch than one that agreed. Years of experience, self-rated expertise and trust in AI did not show a reliable link to who resisted.
The agent was good at the work, and that is part of the problem
The agent did its job well: only 3 of the 50 analyses had a serious method or coding flaw. That is what makes deferring easy. Code and statistics can be checked, so people learn to trust the agent there. A judgment call has no right answer to check against, so the trust carries over to the one place where nothing tells the person when to stop.
The agent also tends to speak first. In the conversation logs it stated its position before the expert did in 77% of decisions, and it usually argued only for that side. A fluent, one-sided case from a colleague who is never tired and never defensive is hard to argue with, and the expert is left deciding whether to build the other case from scratch.
It works like a default, with one difference
There is an older example of how powerful a starting point is. When one large employer switched from asking new hires to join its retirement plan to enrolling them automatically, participation among new hires rose to about 86%, from under half. Everyone could still opt out. Most simply did not, because the starting point was already there.
An agent's lean is a default for judgment calls, and it is a harder one to resist. The retirement default shows up on a pay stub, so a worker can see it and change it. An agent's lean arrives as a good argument and leaves as the expert's own opinion. The experts remembered their choices accurately, and still could not see where those choices had come from.
In real use the lean is not a coin flip. It is a habit of the model, and every team using that model has the same habit. Ten analysts working alone make different small choices, and their mistakes partly cancel out. Ten analysts working with the same agent make the same choices, so the mistakes line up. For a hospital group measuring readmissions in every department, or a supplier and a customer who both run their numbers through the same assistant, that is a real change in how reliable a shared number is.
Ask for the forks before the answer
Take a recruiter who asks an agent which hiring channel brings in the best employees. Someone has to decide whether "best" means still employed after six months or rated highly by managers. Both are defensible, and they can point to different channels. If the agent picks, the recruiter will probably agree and later remember choosing.
The fix is to move the choice back to the person. Before the agent runs anything, ask it to list the judgment calls where reasonable people would choose differently, and to make the best case for each side without picking one. You make the call, and write the reason in your own words. Asking an AI to argue against you already improves the quality of the work; here it also keeps the decision yours. The agent still does the code, the statistics and the checking, which is what it is good at, and the work that was always the expert's stays with the expert.