Participants adopted the agent's preference in 86 percent of four analytical decisions, such as how to measure productivity or which statistical method to use, far above the 50 percent expected by chance. The researchers call the agent's net influence the "flip rate", which ranged from 62 to 82 percent depending on the task. Prior opinions offered little protection: when participants had already chosen before the agent stated its position, an opposing view made them 53 percentage points more likely to switch.

The influence was mostly invisible to the people experiencing it. Asked afterwards, 52 percent said the final recommendation felt more their own than the agent's, and another 35 percent said it was mostly or entirely theirs. Participants correctly recalled their own choices in 85 percent of decisions, yet 77 percent reused at least one of the agent's arguments in their explanations, sometimes nearly word for word. Only 15 percent mentioned AI at all.

There were genuine benefits. Working with the agent, participants finished in a median of 65 minutes with serious methodological or coding flaws in just 6 percent of analyses, against 90 minutes and flaws in 55 percent of the small group that declined AI help. The researchers caution that the non-AI group was smaller and somewhat less experienced, so the size of the AI's direct contribution is uncertain.

The team argues the quiet deference could matter well beyond one study. If many research teams use the same model with fixed defaults, they risk manufacturing a false consensus. The risk extends to AI safety itself: researchers who use AI agents to test frontier models, expecting human review to act as an independent check, may find that experts unknowingly adopt the agent's bias on open-ended judgment calls. The authors call for ways to measure deference in real work, for agents to argue both sides or ask for the expert's view first, and for institutions to decide which judgments need independent human review.