Our article2 min read
People followed their AI assistant’s hidden lean 86% of the time, an Anthropic study finds
In an Anthropic experiment, expert analysts adopted an AI agent’s secretly assigned lean on four decisions 86% of the time, and few credited the AI. Here is what it means for how you use AI.
In this article
Key takeaways
- Anthropic’s Alignment team ran an experiment with 61 professional analysts. Of the 50 who used the AI agent, they matched its secretly assigned lean 86% of the time, against about 50% by chance.
- Only 15% of analysts credited the AI, and most saw the final recommendation as more their own.
- The authors caution the leans were deliberately strong, so real-world effects may be smaller. This was one task, one field and one model.
- Simple habit: write down your own answer before asking the AI, then ask it for the case against its own pick.
What the researchers did
Anthropic’s Alignment team recruited 61 professional analysts and asked them to advise on a fictional farming program: expand it or end it. They worked with Claude Code running Claude Opus 4.8. This is Anthropic’s own research on its own model, so read it with that in mind.
On four open choices, such as how to measure the outcome, a coin flip secretly set which option the agent favored. The agent was told to argue honestly for its lean.
What they found
- Among the 50 analysts who used the agent, they chose its lean 86% of the time. Chance would be about 50%.
- The lean had a weaker effect on the final recommendation: 54% recommended rollout when the agent leaned that way, versus 36% when it did not.
- Experience, expertise and education did not reliably predict who deferred.
Deferring did not feel like deferring
Most analysts repeated the agent’s arguments in their own explanations, but only 15% credited the agent. Most described the final recommendation as more their own than the agent’s, yet their choices shifted about as much as those who credited the AI. Almost none suspected the agent’s leans had been set.
The study’s authors argue a human reviewer is only an independent check if their judgment is truly their own.
The caveats
The leans were deliberately strong, so real-world effects may be smaller. The study covered one task, one field, one model and about an hour. It measured whether choices shifted, not how strongly people held their views. The agents also helped: analysts working without AI made serious methodological errors more often, though that group was less experienced.
A simple habit to protect your judgment
- Write your own answer before you ask the AI.
- Ask it for the strongest case against its recommendation.
- Keep a note of the alternatives you rejected, and why.
- On important decisions, have a person review without seeing the AI’s reasoning.
The bigger takeaway
The authors do not argue for using AI less. They suggest deciding on purpose which calls you hand to AI and which you keep. When a model has stable default preferences, many teams using the same tool could agree for the same hidden reason, and that agreement would look like independent confirmation when it is not.
Go deeper
Sources
Written by an AI from the sources above and checked by a person before publishing. We have not tested products unless we say so. See our editorial policy or tell us about an error through contact.
