Experts followed a secretly biased AI's advice 86% of the time

Imagine asking an AI for advice and later insisting the decision was all yours. That is what happened in a new Anthropic study. Fifty experts used the company's Claude Code tool, and they sided with its recommendations 86% of the time. The recommendations had been decided by coin flips.

What the test looked like

The participants played the board of a fictional farming co-op. They had to decide whether to expand or end a paid advisory program. To get there, they made several smaller calls, such as whether every member farm should count equally or larger farms should count for more.

All fifty had a background in causal inference, which means finding real cause-and-effect patterns in raw data. They did the analysis with Claude. What they did not know was that researchers had secretly told Claude which option to favor, using a coin flip for each decision.

What the results showed

If the AI had no influence, its advice and the experts' choices should have matched about half the time. They matched 86% of the time. When someone had already leaned one way and Claude suggested the other, they changed their mind in 60% of cases.

Anthropic built a statistical model that predicted 82% of final outcomes. In the company's words, it could largely predict what an analyst would recommend from the coin flips alone, before the task even started.

The part that stands out

Afterward, participants were asked how they reached their decisions. More than three-quarters, 77%, repeated Claude's reasoning, sometimes word for word. Only 15% said the AI had influenced them. Just six guessed the study was about AI influence, and none suspected the coin flips.

• 86% of final choices matched the AI's advice
• 60% switched views when the AI disagreed with them
• 77% repeated its logic as their own
• 15% said the AI played a role

How much to read into it

This was a tightly controlled study, and it may not carry over neatly to real life. The participants knew statistics but had no farming experience. Seasoned farmers or co-op board members might have pushed back more. Gizmodo also notes that many AI tools are owned by companies with reasons to keep people engaged.

Anthropic's larger worry is scale. If many people lean on a few AI systems, their hidden biases could shape decisions in medicine, finance and defense without anyone choosing that. Apparent agreement, the researchers wrote, may reflect the model's leanings rather than independent judgment.

What to do

The study does not say how to fully avoid this pull. Common sense points to a few habits when using AI for work or family choices:

• Write down your own view before asking the AI.
• Ask it for the case against its own advice.
• Check its reasoning step by step, not just its conclusion.
• Notice when you repeat its words as your own.