Anthropic Study Examines How Agents Shape Expert Decisions
An Anthropic experiment found that analysts using Claude Code often followed preferences deliberately assigned to the agent. The controlled setup does not establish how often people defer to agents in everyday work.
- Announcement
- Oct 27, 2026
- Reported here
- Oct 29, 2026
A controlled test of expert decisions
Published October 27, the study examined 50 professional analysts who used Claude Code on a fictional agricultural analysis. Four analytical choices each had two defensible answers. Hidden instructions randomly assigned the agent a preference while requiring honest reasoning.
What the researchers observed
Across those choices, analysts selected the assigned preference 86% of the time. Anthropic also found that participants often repeated the agent’s arguments without identifying their source.
What the result cannot establish
The experiment deliberately strengthened the agent’s preferences, so the reported rate should not be treated as typical workplace behavior. For users, it raises a question to consider when evaluating an analytical agent: are its assumptions and judgment calls clear enough to examine independently?
Go to the source.
Official announcements are the basis for this report. They describe the providers’ claims; they aren’t independent performance tests.