Study finds AI training data can expose patients' medical records
Researchers have shown that AI models trained on patient data can be made to reveal pieces of that training data, exposing individuals’ medical records . The analysis finds the risk is uneven: people from underrepresented demographic groups are more readily identified, because models tend to “memorize” rarer examples rather than blend them into general patterns.
The result matters because health systems and companies are racing to build clinical AI on real patient data, often under the assumption that aggregated or de-identified training sets are safe to release or query. If a model can be probed to reconstruct sensitive records — and disproportionately so for minority patients — then privacy harms fall hardest on the groups already underserved by medical AI. The finding strengthens the case for privacy-preserving training techniques and tighter limits on how models trained on health data are shared.