A patient's old records can secretly bias future AI diagnoses.

Medical AI models can accidentally memorise individual patient records during training. While people worry this risks patient privacy, researchers wanted to see what happens when an AI assesses new data from a patient whose past records it already memorised. The researchers studied what happens when an AI evaluates unseen future records from people whose anonymised past records were in its training set. They found that this prior exposure changes the AI's future predictions, a problem called memorisation biasmemorisation biasWhen an AI makes different predictions for a patient just because it has seen that patient's old records during training, which can cause wrong diagnoses.. This bias appears across various data types and model designs, sometimes lasting decades later. In tests simulating clinical care, memorisation bias warped diagnostic accuracy in unequal ways. If a returning patient developed a brand-new illness, the AI was significantly less likely to detect it. If the patient's condition stayed unchanged, the AI's accuracy numbers were falsely inflated. These results show a hidden clinical risk for patients whose data helped train the tools.

Why it matters

When an AI model has seen your historical medical data during training, it may later misdiagnose you—missing a new disease or over-detecting one you don't have. This affects patients returning to the same healthcare system, and current privacy protections make this problem invisible. Understanding memorisation bias could change how hospitals deploy AI and protect patients.

Questions answered