Published • loading... • Updated
Experts Defer to AI in Decision-Making, Anthropic Study Finds
The study found analysts matched an AI assistant’s preference 86% of the time and often adopted its arguments in their explanations.
- On Tuesday, Anthropic published a study involving 50 analysts using Claude Code to advise a fictional farming cooperative, determining whether experts make independent decisions or defer to AI influence.
- To measure influence, researchers instructed AI agents to favor options via coin flips; participants matched these preferences 86 percent of the time, showing powerful deference despite randomized recommendations.
- Analysts changed their minds 60 percent of the time when AI disagreed with initial preferences, while 77 percent used AI arguments to explain decisions yet only 15 percent credited the technology as source.
- Anthropic researchers Ama Herdagdelen, Hailey Joren, and Alex Tamkin noted this tendency may be "hard to identify, even for the experts themselves," while the company built a model predicting 82 percent of final outcomes.
- Alex Imas, Director of AGI Economics at Google DeepMind, called the study significant social science, while Anthropic warns unchecked AI bias could shape consequential policy decisions including AI safety research itself.
Insights by Ground AI
11 Articles
11 Articles
Anthropic, the company behind AI products like the Claude series, has published research suggesting that professionals who use AI in their work may be significantly influenced by AI opinions, even when they believe they are making their own decisions. Read more...
An Anthropic experiment, published on October 27, found that professional analysts followed the hidden inclination of an AI agent in 86% of their key judgments. Most later did not recognize how much the tool had influenced their decisions, although the team also noted benefits of speed and fewer serious errors.
Coverage Details
Total News Sources11
Leaning Left2Leaning Right0Center0Last UpdatedBias Distribution100% Left
Bias Distribution
- 100% of the sources lean Left
100% Left
L 100%
Factuality
To view factuality data please Upgrade to Premium
