What Can Be Done
Concrete paths forward for different audiences. Understanding the problem enables addressing it.
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
- Name pairs, symptom profiles, findings available
- Download research 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.