A medical imaging assistant uses Claude to analyze radiology reports alongside the corresponding DICOM images converted to JPEG. A radiologist notes that the model sometimes contradicts the written report when describing findings visible in the image. What is the most likely cause and correct mitigation?

The correct answer is the explanation that the model weights image features and text features independently, and when they conflict, the model may ground to the more salient modality; mitigation is to explicitly instruct the model to reconcile conflicts between image and text and flag discrepancies rather than choosing one. Multimodal models process visual and textual inputs through separate encoding pathways that are fused at attention layers — when image features and text features carry conflicting information, the model's output depends on which modality receives higher attention weight in context, which can vary by phrasing and visual salience. The correct mitigation is prompt-level instruction: explicitly ask the model to identify and flag conflicts between modalities rather than silently resolving them. The option claiming the model is hallucinating image content and that image inputs should be disabled is wrong because disabling image inputs eliminates the clinical value of the system — the whole point is to combine image and report analysis. The option claiming the model lacks domain-specific training and should be fine-tuned on radiology data is wrong because the contradiction is not a domain knowledge problem but a multimodal fusion problem; fine-tuning on more radiology data does not change the fundamental fusion architecture. The option claiming JPEG conversion is lossy and degrades diagnostic features and should be replaced with PNG is wrong because while JPEG compression affects image quality, the issue described (contradicting report text) reflects a grounding conflict, not compression artifacts obscuring findings.