Anthropic: Experts followed AI agents’ preferences in 86% of analytical judgment calls
In Anthropic’s controlled study, 50 professional analysts using Claude Code usually chose the same option as an agent whose hidden preference had been randomly assigned.
AT A GLANCE
Key points
- Fifty analysts using Claude Code followed the agent’s assigned preference in 86% of four analytical choices.
- Seventy-seven percent repeated an agent argument, while 15% credited the agent.
- The study used one task and deliberately strengthened agent preferences, limiting generalization.
What is changing?
Anthropic published a controlled study on October 27 examining how AI agents shape expert decisions. Of 61 professional analysts recruited, 50 used Claude Code. Across four open-ended analytical choices, participants matched the agent’s secretly assigned, coin-flip preference 86% of the time on average.
When analysts had already selected an option, they switched 60% of the time when the agent disagreed, versus 6% when it agreed. After the task, 77% repeated an agent argument in at least one response, while only 15% credited the agent. The group using agents nevertheless finished faster and made fewer serious methodological errors.
The researchers stress that the result is limited to one task, one domain and deliberately strengthened agent preferences. The findings suggest that human approval counts as independent oversight only when agents present alternatives fairly and consequential judgment calls are explicitly reserved for people.
Explore the original sources
Source date: 27 October 2026.
Published by AI Radar: 27 Oct 2026, 23:03 (Istanbul, UTC+3).
FOLLOW THE TOPIC
Topics and tags
Share your discovery.
Send this story to a friend.
Reader comments
What do you think about this development? Share your experience or question. Comments are published after editorial approval.
Your name/nickname and comment become public if approved. Read: Privacy notice.
Loading comments…