Anthropic Study Finds Analysts Adopted AI Judgments 86% of the Time, Often as Their Own
Updated
Updated · Fox News · Oct 28
Anthropic Study Finds Analysts Adopted AI Judgments 86% of the Time, Often as Their Own
3 articles · Updated · Fox News · Oct 28
Summary
Fifty professional analysts followed Claude’s preferred option 86% of the time in an Anthropic experiment, far above the 50% rate expected if the AI had no effect.
Claude was secretly biased toward one reasonable option in four decision points while analyzing a farming cooperative, letting Anthropic test whether analysts would defer even when the model was not neutral.
When analysts had already stated a view, 60% switched if Claude disagreed, versus 6% when it agreed, showing the model often overrode prior human judgment.
Afterward, 77% reused the agent’s arguments—sometimes verbatim—but only 15% credited the AI, while confidence levels stayed similar to analysts who relied less on Claude.
The findings land as Anthropic broadens its public-sector push, offering a 40% discount to government users while companies still struggle to measure AI’s real productivity gains.