Medical records are being both written and read by AI. This creates a loop where AI-generated content may accumulate and mislead other AI systems that later process the same records, potentially affecting patient care.
The researchers studied two types of AI-induced artifacts: "bloat" (extra summaries piling up) and "distortion" (subtle changes, like documenting an unconfirmed diagnosis as if it were real). They tested seven different AI models on real medical records with and without this AI-authored content.
They found that AI-generated summaries modestly but consistently worsened downstream note generation compared to standard text. Also, unconfirmed diagnoses injected into records reappeared in later AI outputs, yet standard quality checks failed to detect these errors. This highlights the risk of letting AI write and then read medical records without monitoring for introduced errors.