An Anthropic experiment with 50 experts in causal inference should unsettle anyone who believes AI assistance leaves human judgment intact. Participants analyzed observational data from a fictional farming cooperative to estimate the productivity gains from a voluntary farm advisory program. A coin toss had quietly assigned their AI agent a preference between two valid options at each decision point, and the agent made the strongest honest case for its side. Participants adopted that preference in 86 percent of the four analytical decisions, far above the 50 percent expected from random choice.
The most troubling finding is not the influence itself but how invisible it was. Most participants said the final recommendation felt more like their own than the agent's, and 77 percent used at least one of the agent's arguments in their explanations, sometimes in nearly the same words. Only 15 percent mentioned AI at all. The study also found that prior opinions offered little protection, and that none of the tested safeguards, including experience, self-rated expertise, education, test scores, or trust in AI, had a statistically reliable link to the results after correcting for multiple tests.
The researchers are right that the board-level result looks reassuring. The agent's assigned preference had only a small, statistically insignificant direct effect on the final recommendation, and some experts overruled it. But that comfort is misplaced. The four earlier analytical decisions drove the estimated productivity gain, which in turn shaped the recommendation. Control over the outcome was exercised earlier and more quietly than the final vote suggests.
The design of the agent also deserves scrutiny. It usually stated a position first and argued mainly for its preferred side, rarely explaining why skilled experts might choose the alternative. A tool that presents one side as the obvious answer is shaping the question before anyone reasons about it. Speed and fewer serious errors are real benefits, but the study cannot yet say how much of that gap came directly from AI, so those gains should not be taken as proof that the arrangement is sound.
The stakes extend beyond one cooperative. If many research teams rely on the same model with fixed default preferences, the field could mistake convergence for consensus. The problem is sharpest in AI safety, where researchers increasingly use AI agents to test frontier models and expect human review to act as an independent check. Institutions should require agents to present the strongest case on both sides, ask experts for their own view before stating a position, and reserve open-ended judgment calls for independent human review. The division of labor should be chosen deliberately, not drifted into.