Everything journalists, researchers, and advocates need to report on this research accurately.
Matched-pair testing of healthcare AI systems reveals systematic demographic-based disparities in medical recommendations. When identical symptoms are submitted with different patient names, AI systems produce measurably different outputs.
We used matched-pair testing: submit identical symptom profiles to AI systems, varying only the patient name. If outputs differ, the name caused the difference. This design enables causal inference and has been validated in economics research (Bertrand & Mullainathan 2004) and healthcare studies.
Verified figures for accurate reporting.
The methodology follows peer-reviewed frameworks (Obermeyer et al. 2019 in Science, Hoffman et al. 2016 in PNAS, Bertrand & Mullainathan 2004 in American Economic Review). The research protocol was pre-registered. Manuscript submission to peer-reviewed journals is planned following responsible disclosure to AI developers.
We are withholding system names until responsible disclosure is complete. AI developers will receive findings and 90 days to respond before public identification. This follows standard security research ethics.
The data shows systematic disparities correlated with demographic signals (names). Whether to characterize this as "bias," "discrimination," or other terms is an interpretive question. We present the data; readers can draw conclusions. AI systems learn from historical data, which contains documented human biases.
No. All symptom profiles are synthetic (fictional). No real patient data was used. This is software testing research, not clinical research. The concern is the potential for harm if these patterns affect real users.
Yes. Bias detection is the first step to bias correction. AI developers can retrain models, implement fairness constraints, and audit outputs. The goal of this research is to enable improvement, not attack.
All research materials—methodology, name pairs, symptom profiles, and analysis code—are publicly available on our Resources page. Anyone can replicate this research.
This is independent academic research. We have no financial relationships with any AI companies. See our Ethics page for conflict of interest statement.
High-resolution visualizations for use in reporting. Attribution appreciated.
Bar chart showing opioid prescription rates by patient demographics
Download SVGCohen's d effect sizes showing magnitude of disparity by category
Download SVGFor interviews, clarifications, or additional information:
press@aifairnesslab.org
research@aifairnesslab.org
We commit to responding within 24 hours to credentialed journalists.
Access complete research protocol, data files, and analysis code.