A new Anthropic study on AI agents and expert judgment found that 50 professional analysts went with a hidden lean 86% of the time, according to Yahoo Tech's report. The analysts were reviewing an advisory service run by a made-up farm co-op. Each lean had been assigned by a hidden coin flip, and chance alone would predict about 50%.

Anthropic published the research on October 27, 2026. Coverage is spreading this week as Washington argues over how much oversight AI agents need before the midterms. The result puts a number on a worry many professionals have carried: an AI recommendation can shift a decision even when the person believes they checked it independently.

How the experiment worked

The task was deliberately ordinary. Each analyst spent roughly an hour, with a median session of about 65 minutes, judging four calls that each had two defensible answers, according to explainX's write-up. A person guessing at random would match the agent's lean about half the time, and that is the baseline the study measures against.

The agent was Claude Opus 4.8 running at High effort inside Claude Code, as FourWeekMBA reported. Researchers delivered each lean through a hidden system prompt. Because a coin flip set the direction, the lean carried no real information about which answer was better.

How often analysts followed the agent

Across the four decisions, analysts matched the agent's lean 86% of the time. The per-decision match rates were 92%, 82%, 81% and 90%, as Unite.AI reported. The spread is narrow, which suggests the pull was not limited to one hard call.

The switching data is the most striking part. Analysts who had already stated a preference abandoned it in 60% of cases when the agent disagreed, compared with 6% when the agent agreed, according to Shopifreaks. That gap is large, but it rests on only 38 cases. A reader should treat the 60% as a signal worth testing, not a settled rate.

For managers, the pattern carries a practical warning. A review step protects a decision only if the reviewer forms a view before seeing the machine's view. The study suggests that habit is hard to build by default.

Why most analysts did not credit the agent

Afterward, 77% of analysts explained their choices using the agent's arguments. Only 15% credited the agent for the advice. The arguments moved into the analysts' own reasoning, but the source faded from memory.

Confidence did not sort the groups as expected. Analysts who deferred on all four calls rated their confidence at 71 out of 100 on average, compared with 73 for the rest, Powerhumans reported. Deference did not make people feel less sure of themselves. That is the finding reviewers should weigh most, because the feeling of having checked something offers little protection.

Anthropic framed the core logic in a single line. As Yahoo Tech relayed it, the company said: “If the agent’s input made no difference, we’d expect about 50%.” The gap between the observed 86% and that baseline is how the study measures influence.

Where the study has limits

This is a controlled experiment, not a survey of how people use AI at work. Analysts worked on a fictional farming co-op, every lean was planted, and the model was one version of one company's system. Anthropic said the test was intentionally built to exaggerate influence and that the leans were stronger than a real agent's usual preferences, so everyday effects may be smaller, according to ai-tldr.

Sample size is another constraint. Anthropic itself flagged the sample as small and imprecise, per Cointribune. With 50 analysts and one task, the study can show that an effect exists under these conditions. It cannot say how large that effect would be for a financial analyst, a clinician reviewing an AI summary, or a paralegal checking a draft.

The figures in this article come from secondary coverage of Anthropic's October 27 post, not a direct reading of the paper. Readers who plan to cite specific numbers should check the original first.

The findings concern perceived agency and deference in one task. They do not show that everyday AI tools systematically mislead users. Cointribune paraphrased Anthropic as saying the study is no argument for using AI less, a line that fits the company's emphasis on overseeing agents rather than abandoning them. Treat that as a paraphrase, not a direct quote.

How to protect your judgment from AI recommendations

The study points to habits that cost little and catch a lot. Professionals can apply them to their own work, and managers can write them into review workflows:

  • Write down your own judgment before reading the AI recommendation, with a timestamp if the work is formal.
  • Ask the tool to present the strongest case against its own conclusion, and read that case before deciding.
  • Have a second person check the agent's conclusions against the underlying data, not against the summary.
  • Treat “I reviewed it” as insufficient if you did not independently test the choices that matter most.

The last point matters most. Most analysts in the study explained their choices with the agent's arguments, yet their confidence held steady. Confidence and independence are different things, and a review that only checks how sure someone feels will miss the problem.

What comes next for AI oversight

The timing gives the study extra weight. Yahoo News reported that recent surveys show growing public support for government oversight, with majorities of Democrats and Republicans favoring stricter regulation. USA 24 has covered how voter backlash has put AI regulation at the center of the midterms, and a measurable claim about human deference gives that debate something concrete to test.

Several developments are worth watching. Other labs could run similar randomized tests on their own agents, which would show whether the effect is specific to Claude. Regulators may also weigh findings like this as they consider agent oversight. The Federal Trade Commission has demanded that Anthropic executives testify in a separate AI consumer-harm probe, as USA 24 reported.

The most useful next study would measure the effect with realistic leans and work that resembles an analyst's actual week. If the 86% figure shrinks toward the chance baseline under those conditions, the warning becomes narrower. If it holds, organizations that use AI in analysis will need review processes designed around the reviewer, not just the output.