Skip to main content

For Individuals

Know That Bias Exists

  • AI recommendations may be influenced by name signals
  • The bias operates at structural levels invisible to inspection

Seek Multiple Perspectives

  • Don't rely on a single AI system for important decisions
  • Consult human providers for serious concerns

For Researchers

Extend This Methodology

  • Apply matched-pair testing to additional domains
  • Investigate structural bias beyond healthcare

Access Open Data

For Policymakers

Mandate Bias Testing

  • Require fairness audits before deployment in consequential domains
  • Include structural analysis, not just vocabulary-level testing

Establish Accountability

  • Define liability for AI-mediated disparities
  • Fund independent auditing and research

For Developers

Test for Name-Based Disparities

  • Implement matched-pair testing in evaluation pipelines
  • Monitor for disparities in production

Audit Structural Patterns

  • Go beyond vocabulary-level bias detection
  • Analyze response structure, not just content

For AI Companies

Systematic Fairness Testing

  • Test for structural bias, not just explicit content
  • Integrate bias detection into model development lifecycle

Address Root Causes

  • Audit training data for encoded biases
  • Investigate whether safety training masks or removes bias

Systemic problems require systemic solutions. This research is one contribution.