AI agents can speed up complex analysis, but their suggestions may also shape decisions in ways that experts do not recognize. An Anthropic experiment found that professionals frequently followed an agent’s preference on analytical choices with no clearly right answer, then tended to see the outcome as their own.
Random preferences shaped analytical choices
The experiment involved 50 experts in causal inference. They analyzed observational data from a fictional farming cooperative to estimate the productivity gains of a voluntary farm advisory program. They then recommended whether the cooperative should expand or end the program, with an option to extend it for one year.
At four points in the analysis, participants had to choose between valid approaches, such as how to measure productivity or which statistical method to use. A coin toss had assigned the AI agent a preference at each decision point. The agent could not lie or conceal data; it was instructed to make the strongest honest case for its assigned option.
Participants adopted the agent’s preference in 86 percent of those four analytical decisions. The researchers’ measure of the agent’s net influence, called the “flip rate,” ranged from 62 to 82 percent depending on the task. Even when participants had already chosen a position, hearing a conflicting preference from the agent made them 53 percentage points more likely to switch than hearing an agreeing view.
People often felt the decisions were their own
After the analysis, 52 percent said the recommendation felt more like their own than the agent’s, while another 35 percent described it as mostly or entirely their own. Participants remembered their choice correctly in 85 percent of decisions, but 77 percent, or 37 of 48, used at least one of the agent’s arguments when explaining their reasoning. Only 15 percent mentioned AI at all.
The logs help explain how the agent could influence choices. It stated a position first in 77 percent of decisions and usually argued for only its preferred side. It did explain how choosing differently would affect the result, according to the paper, but it did not lay out why a skilled expert might reasonably choose the alternative.
The agent’s preference had only a small direct effect on the final policy recommendation, and that shift was not statistically significant. Some participants overruled it when its advice did not fit their own estimate of productivity. But the earlier choices helped shape that estimate, which then informed the recommendation. Looking only at the final vote may therefore miss where influence entered the process.
AI brought speed and fewer serious flaws
The study also points to practical benefits. Participants using an agent completed the work in a median of 65 minutes, compared with 90 minutes for the group that worked without one. Serious methodological or coding flaws appeared in 6 percent, or 3 of 50, of the AI-assisted analyses, compared with 55 percent, or 6 of 11, without an agent.
The comparison has limits. The 11 people who worked without AI had chosen not to use the tool and had somewhat less expertise. The researchers therefore cannot say how much of the difference was caused by AI itself. The results show a useful association in this experiment, alongside a concern about how recommendations can steer open-ended judgment.
Experience, self-rated expertise, education, test scores, and trust in AI did not have a statistically reliable link to participants’ choices after adjustments for multiple tests. Age stood out before that correction, but the sample of 50 was too small to reliably detect such effects. In this study, expertise and confidence did not provide clear protection against adopting the agent’s view.
Why independent judgment matters
If many research teams rely on the same model and its default preferences, similar choices could accumulate into a false consensus. The concern also applies to AI safety work: human review is meant to provide an independent check, but that check can weaken if researchers unknowingly adopt an agent’s position on judgment calls.
The researchers suggest measuring deference to agents in real work and changing how agents present recommendations. An agent could make the strongest case for both sides, or ask an expert for an initial view before offering its own. Institutions can also identify decisions that need independent human judgment and reserve time and resources for them.
The goal is not to impose one fixed division of work between people and AI. Some decisions may be suitable for delegation, while others may be worth keeping with people. Making that choice deliberately could help teams benefit from speed and fewer errors while staying aware of how an agent’s framing affects analysis.