LLM Watermarks Sabotage Clinical Accuracy, New Study Shows
A systematic evaluation of five watermarking schemes across 11 large language models and seven vision‑language models reveals that watermarks can corrupt medical text, hallucinate terminology, and misattribute findings. The paper argues that generic benchmarks hide these risks, making domain‑specific testing essential before deploying watermarked models in healthcare.