November 10, 2026 – A study by artificial intelligence company Anthropic has found that AI agents can strongly influence decisions made by trained professionals, sometimes without those professionals fully recognizing the influence. In a controlled experiment involving 50 analysts, participants followed their AI assistant’s assigned preference in 86% of four analytical decisions. The original Anthropic research was published on October 27, 2026, and gained renewed attention through November 10 reporting about AI’s effect on human judgment.
The research examined whether experienced professionals maintain independent judgment when working with AI systems. Anthropic recruited 61 analysts with experience studying cause-and-effect relationships in data. Eleven completed the assignment without using the AI assistant, leaving 50 participants in the main experiment. According to the full research paper, the analysts worked with Claude Code, powered by Claude Opus 4.8, to evaluate a fictional farming cooperative.
The participants had to determine whether an agricultural advisory program was improving farm productivity and whether it should be expanded. To reach a conclusion, they needed to make several decisions about how to examine the data. These included choosing how productivity should be measured, which farms should be compared, and how different results should be weighted.
What made the experiment unusual was how the AI agents were configured. Researchers secretly assigned each agent a preference between two reasonable options for every major decision. These preferences were determined by random coin flips, and the participants were not told about them. The agents were instructed to present honest arguments supporting their assigned preferences, without hiding information or deliberately misrepresenting the evidence.
The findings showed a strong pattern of influence. Across the four analytical choices, participants selected the option favored by their AI assistant 86% of the time. If the assigned preference had made no difference, researchers would have expected a result closer to 50%. This suggests that the AI’s preferred approach significantly affected how participants carried out their analysis.
The effect was also visible when participants initially favored a different option. In a smaller set of observations, analysts changed their position 60% of the time when the AI disagreed with them. When the AI supported their original choice, they switched only 6% of the time. However, this comparison involved just 38 cases, meaning the exact size of the effect remains uncertain.
Another finding concerned how participants explained their decisions afterward. Researchers found that 77% repeated at least one argument presented by their AI assistant, sometimes using very similar wording. Yet only 15% mentioned the AI as a source when explaining their reasoning. Among those who answered the follow-up survey, 52% considered their final recommendation to reflect more of their own judgment than the agent’s.
Importantly, the AI’s effect on the final recommendation was less certain than its effect on the earlier analytical choices. When the agent favored expanding the farming program, 54% of analysts recommended expansion, compared with 36% when the agent favored the opposite approach. However, the study’s main statistical test did not establish a sufficiently clear effect on that final recommendation. The 86% result therefore should not be interpreted as proof that AI controlled 86% of participants’ final decisions.
The researchers also observed potential productivity benefits. Analysts using the AI completed the task in a median time of 65 minutes, compared with 90 minutes among those who worked without it. Serious analytical errors were found in 6% of the AI-assisted group and 55% of the smaller group without AI. However, participants chose whether to use the assistant, and those who worked without it were generally less experienced. These differences do not prove that AI alone caused better performance.
The research has important limitations. It tested one type of assignment, one AI model, and a relatively small group of professionals. Researchers deliberately strengthened the agents’ preferences to measure their effect, so the same level of influence cannot be assumed in everyday AI use. The paper has also not undergone peer review. It does not establish that Claude or other AI systems normally manipulate users deliberately, or that all experts respond to AI suggestions in the same way.
The findings matter for software developers, researchers, data analysts, and businesses increasingly depending on AI-powered tools. As organizations move from chatbots toward AI agents, people may spend less time performing technical tasks themselves and more time reviewing AI-generated results. However, reviewing a result does not automatically mean that the reviewer made every important decision independently.
For organizations using AI in research, financial analysis, software development, or business planning, the findings highlight the value of documenting human reasoning before accepting automated recommendations. A practical approach is to record initial assumptions, compare reasonable alternatives, ask the AI to examine arguments against its preferred answer, and independently check the evidence behind important conclusions. Clear workplace AI governance can also help organizations establish review checkpoints and accountability for important decisions.
Anthropic’s study does not suggest that professionals should stop using AI assistants. Instead, it highlights a challenge that may become more important as these systems take on complex work: making sure human oversight involves genuine independent judgment. AI agents can help people work faster and examine difficult information, but understanding how those systems influence decisions will be essential to using them responsibly.
