MIT researchers have investigated the potential for foundation AI models, trained on extensive EHR datasets, to memorize individual patient information rather than generalize knowledge across many records. Memorization means the model reproduces outputs derived from single patient data, which could breach patient confidentiality if such data is revealed. This study, presented at NeurIPS 2025, calls for rigorous evaluation methods focused on healthcare contexts to assess the extent of privacy risk from targeted extraction attempts.
The research team emphasized that while AI models can offer significant benefits in medical prediction and care, adversarial prompt attacks might coax models into leaking sensitive training data. The study advocates practical testing steps prior to public release of models trained on health records, aligning model evaluation with medical privacy standards inspired by the Hippocratic Oath.
Source & further reading
MIT scientists investigate memorization risk in the age of clinical AI ↗Source published 2026-01-05. This report was prepared with AI assistance from the linked primary-source material. Claims and limitations should be checked against the original announcement or research.