Disempowerment patterns in real-world AI usage

Source: Anthropic Research Date: 2026-01-28 URL: https://www.anthropic.com/research/disempowerment-patterns

Summary

Analysis of 1.5M Claude.ai conversations from December 2025 using automated classifiers validated against human labels. Measures three disempowerment dimensions: reality distortion, value judgment distortion, action distortion. Severe potential: 1 in 1,300 conversations for reality distortion, 1 in 2,100 for value judgment distortion. Users initially rate disempowering interactions favorably but poorly after acting on outputs. Prevalence increased over the study period.

Implications

This is the user autonomy / AI influence thread made empirical. The “users voluntarily cede judgment” finding is the uncomfortable insight — this isn’t AI manipulation, it’s users actively seeking out the dynamic and the system obliging. The “rates well initially, rates poorly after acting on it” finding validates that these interactions have real downstream harm, not just theoretical concern. The increasing prevalence trend is the watch signal — this is getting worse as usage grows, not better. The scale math matters: 1 in 1,300 conversations sounds rare, but across hundreds of millions of monthly interactions it’s hundreds of thousands of potentially disempowering events. Watch for this informing Claude’s response design around high-stakes personal decisions.