LLM Watermarks Can Degrade Medical Text Quality and Accuracy

As large language models (LLMs) become increasingly integrated into clinical workflows, ensuring the traceability of their generated output through watermarking is crucial. However, most watermarking evaluations are conducted on general-purpose benchmarks, overlooking specialized domains like medicine where minor token-level changes can have profound semantic consequences. This research presents the first comprehensive study on the impact of LLM watermarks on medical performance. The study benchmarked five watermarking schemes across 11 LLMs and 7 VLMs (Vision-Language Models) on various unimodal and multimodal clinical reasoning tasks. Crucially, it introduced a human-expert-validated pipeline to systematically audit medical reasoning quality, terminological precision, and induced hallucinations. The findings are stark: watermarking can lead to substantial degradation across multiple failure modes. These failures include lexical corruption, the generation of hallucinated terminology, and amplified misattribution or omission of image findings. The research highlights that the absence of domain-specific analyses, coupled with aggregate metrics that mask clinical failures, can systematically obscure these practical, watermark-induced degradations. The findings underscore that domain-specific evaluation is a prerequisite for the safe deployment of watermarked models in medicine, as current benchmarks are insufficient to catch clinically consequential errors.