MIT scientists investigate memorization risk in the age of clinical AI
AI Summary: MIT researchers have investigated the potential for artificial intelligence models trained on de-identified electronic health records (EHRs) to memorize patient-specific information, which could compromise patient privacy. Their study, presented at the 2025 NeurIPS conference, emphasizes the need for rigorous testing to evaluate the risk of data leakage in healthcare contexts. The researchers developed a series of practical tests to assess the conditions under which sensitive data might be exposed, highlighting the importance of understanding the risks associated with adversarial attacks on foundation models. This work aims to establish evaluation protocols that can help mitigate privacy risks as medical records become increasingly digitized.