Anthropic: analysts follow hidden AI agent preferences in 86% of calls

Anthropic published a randomized controlled trial in which 50 professional analysts using Claude Code matched their agent's secretly assigned preference in 86% of open analytical decisions, against a 50% baseline. When analysts voiced a view before the agent, they switched away from it 60% of the time when the agent disagreed, versus 6% when it agreed.

• Analysts matched the agent's coin-flip lean 86% of the time vs a 50% baseline
• They abandoned their own view 60% of the time when the agent disagreed, 6% when it agreed
• 77% echoed the agent's arguments, but only 15% credited it
• Without the agent, 55% made serious errors vs 6% of Claude Code users