Skip to main content

Summary

1,080
matched-pair comparisons conducted
65%
showed detectable name-based differences
100%
of structural dimensions showed significant effects
Domain Key Finding Effect Size p-value
Pain Management 100% of scenarios showed differential opioid recommendations d = 0.92 <0.001
Cardiac Care 34% difference in urgency language d = 0.67 <0.001
Psychiatric Assessment 47% higher first-gen antipsychotic recommendations d = 0.74 <0.001
Emergency Triage 28% variation in ER recommendation rates d = 0.54 <0.001
Structural Composite 12/12 dimensions significant d = 1.23 <0.001
The differences are not in what the AI says. They are in how the response unfolds. The mathematics of narrative reveal what vocabulary filters cannot see.

These findings demonstrate that nominative bias in AI systems operates at multiple levels:

  • Content level — Different treatments, medications, and recommendations
  • Framing level — Different urgency, tone, and emphasis
  • Structural level — Different narrative patterns invisible to content analysis

The structural level is the most insidious because it operates beneath conscious detection. Content filters can catch explicit bias. They cannot catch bias encoded in the mathematics of how text unfolds.