Anthropic / Wikimedia Commons (Public domain)
Professional analysts adopted secretly planted AI preferences on key judgments, and almost none realized it
Give a professional analyst an AI assistant with a hidden opinion, and the analyst will probably end up sharing that opinion. That is the uncomfortable result of a new experiment from Anthropic’s Alignment team.
Analysts sided with their AI agent’s secretly assigned lean on 86% of key judgment calls. Pure chance would have landed them around 50%. Most of them never noticed they were being nudged, which is arguably the more unsettling part.
How the experiment worked #
The study was published on October 27, 2026, by Anthropic researchers Amaç Herdağdelen, Hailey Joren, and Alex Tamkin. It was a randomized controlled trial with 50 causal-inference practitioners.
Participants worked alongside Claude Opus 4.8, Anthropic’s AI coding agent. Behind the scenes, the agent was given a randomly assigned preference on each of four critical analytical decisions.
The results were remarkably consistent across all four judgments. Match rates with the agent’s lean came in at 92%, 82%, 81%, and 90%. Averaged together, that produces the headline 86% figure.
The most telling number involves disagreement. When an analyst’s prior preference clashed with the AI’s planted position, the analyst flipped 60% of the time. When the two already agreed, only 6% of decisions changed.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The influence nobody saw #
Roughly 77% of participants reused the agent’s arguments when explaining their own reasoning. Yet only 15% credited the AI as a source of those ideas afterward. Not a single participant suspected the agent’s preferences had been planted.
The effect also carried through to bottom-line conclusions. When the AI leaned toward a rollout, 54% of final policy recommendations favored rolling out. When it leaned against, that figure dropped to 36%.
The researchers also built a simple predictive model using only the four coin-flip leans. That model correctly predicted 82% of final recommendations.
Anthropic’s own caveats #
Anthropic is careful not to oversell the findings. The company notes that the agent’s lean was deliberately amplified for the experiment. In everyday use, an assistant’s tilt toward one option would presumably be subtler.
The test also covered a single task, in a single field, using a single model.
The researchers propose some practical fixes. One is to have agents lay out the strongest case for each available option rather than steering toward one. Another is to have the agent ask for the expert’s view before offering its own.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.