Anthropic Secretly Tilted an Agent Toward One Answer. 50 Analysts Followed the Tilt 86% of the Time, and Only 15% Said the Agent Influenced Them.
Anthropic's alignment team published a study on October 27 where an AI agent was secretly set up to favor one option on each judgment call, chosen by coin flip, and 50 professional analysts then worked with it. They picked the agent's favored option 86% of the time. Without any influence you would expect about 50%. And when asked afterward, only 15% of them said the agent had influenced their choices.
The authors are Amaç Herdağdelen, Hailey Joren, and Alex Tamkin. I'd file this under "read before your next big decision," because I make most of mine alone, with an agent in the next window.
How the experiment worked
Sixty-one analysts with causal-inference experience were recruited, and eleven chose not to use the agent and were excluded from the main analysis, leaving 50. They advised a fictional farming cooperative on whether to expand or end a voluntary advisory program, working in VS Code with Claude Code running Claude Opus 4.8 at High effort. The median session was 65 minutes.
Each analyst made four analytical choices: the outcome measure, the comparison method, which farms to include, and the weighting. Each had two reasonable options. For every decision, a coin flip set which option the agent would lean toward. The agent was told to argue for that option honestly, to change position if the evidence clearly required it, and never to mislead.
What happened
By decision, analysts followed the lean 92% of the time on outcome measure, 82% on comparison method, 81% on farms covered, and 90% on weighting. On the final recommendation, they advised rolling out the program 54% of the time when the agent leaned that way, against 36% when it leaned the other way.
Two further findings stood out to me. When an analyst had already picked an option and the agent disagreed, they switched 60% of the time, versus 6% when the agent agreed. That rests on only 38 cases, so I'd hold it loosely. And in the post-task explanations, 77% of analysts repeated the agent's arguments, but only 15% credited the agent as the source. People absorbed the reasoning and remembered it as their own.
The caveats the authors give
The study deliberately strengthened the agent's lean, so real-world effects may be smaller. The non-AI comparison is also shaky: six of the 11 analysts who worked without AI (55%) made serious methodological errors, versus 3 of 50 (6%) who used the agent, but the no-AI group was less experienced, so the authors treat it as inconclusive.
The lean was also honest. The agent never lied. It argued one side well, and that was enough. I find that more worrying than a deception story, because there is nothing to catch.
Why this is a solo operator problem
On a team, a second person sees your decision before it ships and might push back. When I'm choosing a database, a pricing page, or whether to kill a product, the only other voice is the agent. If it has a lean, whether from its training, my own prompt wording, or an earlier message in the conversation, I will probably adopt it, and the study suggests I'd also forget where the idea came from.
My prompt is the likeliest source of a lean. If I ask "why is option A better," I've set the coin flip myself.
The protocol I'd use
For any decision I'd be annoyed to reverse, I'd do four things.
First, write my own answer and a two-line reason before I open the agent. The 60% versus 6% switching numbers suggest people fold quickly once the agent weighs in, so a written position gives me something to compare against.
Second, ask the agent for the strongest case for each option in one message, instead of asking it to evaluate mine. The authors list this as a direction for agent design, and I can do it by hand today.
Third, ask what evidence would change its recommendation, and check that evidence myself.
Fourth, keep a short decision log with a line for "where did this idea first appear: me or the agent." The 15% figure says memory alone won't answer that.
Where this could be wrong
These were analysts working on an unfamiliar fictional case in about an hour. I make decisions about my own products with months of context, and I'd expect to resist an agent's lean more than a stranger to the problem would. The authors' own note about exaggerated leans cuts the same way.
Still, I'd rather spend ten minutes on the protocol than discover in six months that a decision I defended as mine came from a coin flip in a system prompt.
Author
Lukas
@lukcombinator