Anthropic conducted a study with 50 experts, to whom a system with randomly assigned preferences was presented. In situations where a judgment was required, participants followed its position in some cases up to 86% of the time.
Most experts nonetheless considered the conclusions to be their own. This is important for workflows where professionals use AI as an advisor: the system's influence may remain unnoticed even when the person believes the decision is independent.
The data are published by The Decoder, citing the Anthropic study. The available package contains only metadata and a brief description, so details of the methodology, participant makeup, and experiment conditions require further verification.
editorial commentary
Why it matters
Expected consequence — increased interest in checks of the independence of expert decisions when using AI. The next observable signal will be the publication of the full methodology and results of the study. Significant uncertainty is tied to the fact that only a brief description without primary materials is currently available.