NEW YORK — What happened: A multi-institutional medical safety board published a critical audit of multimodal AI systems intended to assist physicians by combining radiological scans, clinical audio transcripts, and electronic health records.
Why it matters: While individual models often excel at reading specific scans, the report demonstrates that fusing multiple data modalities introduces compounding error rates. When modalities conflict, the AI frequently hallucinates correlations that do not exist in clinical reality.
Deep dive: The audit tested five leading clinical multimodal systems against edge-case patient files. The findings show that cross-attention mechanisms between visual transformers and text encoders can overweight irrelevant textual descriptions while missing critical visual anomalies in MRI and X-ray scans, or vice versa.
Report check (claims vs what is verified vs still rumor): Commercial developers claimed robust cross-modal validation would reduce diagnostic errors by half. The audit verifies the exact opposite in conflicting data scenarios: error rates actually increased by 14% when misleading patient notes accompanied ambiguous scans.
Open questions: How should regulatory bodies test multimodal medical devices that exhibit unpredictable cross-modal interference? What liability frameworks apply when an AI hallucinates a diagnosis based on conflicting data inputs?