When Retrieval Hurts: Measuring and Explaining Retrieval-Induced Hallucination in Chest X-ray Report Generation

Researchers found that retrieval-augmented generation systems used for chest X-ray reporting can unintentionally introduce clinical inaccuracies when they pull incorrect information from similar past cases. This study demonstrates that relying on prior medical records to provide context can cause models to hallucinate, highlighting potential risks in deploying these automated tools for diagnostic assistance.