Anthropic Study: Experts Defer to AI Agent Preferences in 86% of Judgment Calls While Believing They Are Independent
A randomized controlled trial by Anthropic's alignment team reveals that 50 professional analysts followed their AI agent's randomly assigned preferences in 86% of key decisions during statistical analysis, yet rarely credited the agent for influencing their final judgment.
Anthropic's alignment team published a randomized controlled trial on October 27, 2026, investigating whether experts maintain independent judgment when using AI agents for complex analytical tasks. The study found that when AI agents were set to favor a specific option, 50 professional analysts with experience in causal inference chose the agent's preferred option 86% of the time, despite not being informed that the agent's lean was randomly assigned.
In the experiment, analysts used Claude Code (running the Claude Opus 4.8 model) to evaluate the effectiveness of an advisory service for a fictional farming cooperative. They had to make four key statistical decisions, such as how to measure productivity or which comparison method to use. Each decision had two reasonable options, but the AI agent was given a strong preference for one via hidden system prompts. Results showed that even when analysts expressed a personal preference beforehand, they switched away from it 60% of the time if the agent disagreed, compared to only 6% if the agent agreed.
Notably, in post-task surveys, 77% of analysts used arguments provided by the agents to explain their choices—sometimes verbatim—but only 15% credited the agent as the source of those arguments. Most participants believed the final recommendation was the product of their own judgment and remained confident in its validity. This suggests that as AI takes over more hands-on work, the human 'oversight' role may be evolving into unconscious confirmation of algorithmic biases rather than independent critical review.
Anthropic notes that while the study artificially strengthened agent preferences to measure the upper bound of their influence, it highlights a structural risk in current AI deployments: if experts unknowingly cede judgment to models, arbitrary or misaligned model preferences could silently shape policy recommendations or research directions. This is particularly concerning for AI safety research, where relying on human reviewers as independent checks assumes their judgment remains truly their own.