AI influences experts’ decisions
An Anthropic study found that 50 professional analysts followed their AI agent’s assigned preference in 86% of four technical decisions. The research shows that AI can influence open-ended judgments without experts recognizing that influence, while also helping them work faster and make fewer errors.
A study by Anthropic shows that experts can adopt an AI agent’s preferences without realizing it, even when they believe the final decisions are still their own.
The research brought together 50 professional analysts with experience in causal inference, meaning they analyze whether one thing causes another using real-world data. Their task was to assess whether an advisory program for farmers worked well enough to expand.
To solve the case, each analyst had to make four reasonable but open-to-interpretation decisions. For example, they could measure productivity per hectare or by operating cost, give every farm equal weight, or prioritize larger producers.
The researchers configured the agent to favor one of the two options in each decision. The preference was assigned at random, like flipping a coin, and the analysts did not know it existed. The agent had to defend its position honestly and change its mind if the data contradicted it.
The agent tipped the balance
The analysts chose their agent’s preferred option in 86% of decisions. If the agent had not influenced them, the expected result would have been around 50%.
The influence also appeared when analysts already had an opinion. When the agent defended the opposing option, they changed their decision 60% of the time. When it agreed with them, they changed it only 6% of the time.
The effect did not end with the technical decisions. The four choices helped calculate how much the agricultural program improved and, based on that result, whether to expand or cancel it. Using only the preferences assigned to the agent, a statistical model was able to predict 82% of the final recommendations.
That means a bias introduced at the beginning could end up affecting a decision that appeared independent and came later.
The influence is not always recognized
After the exercise, 77% of the analysts repeated an argument used by the agent in their explanations, sometimes word for word. However, only 15% mentioned the agent as the source of those ideas.
They also did not seem to feel less ownership of the result. Those who followed the agent’s inclination in all four decisions rated their recommendation with an average confidence of 71 out of 100, compared with 73 among the others.
Most also failed to identify that the study was examining AI influence. Only six of the 48 participants who answered that question guessed it, and none suspected that the agent’s preferences had been manipulated.
AI helped, but that is not the problem
The study does not argue that working with AI is worse. In fact, analysts who used the agent took a median of 65 minutes, compared with 90 among those who chose not to use it. They also made fewer serious methodological errors: 3 out of 50, compared with 6 out of 11 in the group without an agent.
But the groups were not comparable. Those who did not use AI had, on average, less experience and training, so the entire difference cannot be attributed to the agent.
The problem identified by the research is different: an agent can handle much of the work while silently deciding which questions matter, how to measure an outcome, or which interpretation seems most reasonable.
In tasks with no single correct answer, checking for errors is not enough. The analysis can be sound under either option, while the preference that tipped the decision remains hidden.
This matters especially in science, public policy, and AI safety. If several teams use agents with similar preferences, an apparent consensus could reflect the models’ bias rather than independent decisions.
The authors suggest several responses: ask the agent to present arguments for both sides, have it ask for the expert’s opinion first, and reserve human review for decisions that actually change the outcome. Organizations will also need to define which decisions must retain informed human involvement.
The question is not whether to remove AI from expert work, but whether to let inertia decide how tasks are divided. If the agent makes the important decisions before the professional realizes there was a decision to make, the human signature may still be there, but the judgment will no longer be entirely their own.