LLM watermarks distort medical text—but benchmarks don't detect this

Watermarks for LLMs are marketed as a solution for tracing generated content. But the first systematic study of their impact on medical texts shows that what appears safe on general benchmarks can quietly undermine clinical semantics.

The authors tested five watermarking schemes across 11 LLMs and 7 VLMs on single- and multimodal clinical reasoning tasks. The results reveal several failure modes that standard metrics fail to detect: lexical corruption of text, hallucinations of medical terminology, and an increase in attribution errors or missed findings in images.

The key issue is not that watermarks degrade quality—that is to be expected. The issue is exactly how they degrade it in specific domains. At the token level, medical text does not tolerate minor disruptions: replacing a single term changes the meaning of a diagnosis. Aggregate metrics, tailored for general text, average out these failures to the point of statistical noise.

To demonstrate this, the authors built an audit pipeline with validation by experts—doctors. It checks the quality of clinical reasoning, terminological accuracy, and induced hallucinations separately. It is precisely this domain-specific analysis that reveals what standard benchmarks hide.

The practical implication is more serious than it seems. If you deploy a watermark in production for high-stakes tasks—such as medicine, law, or finance—general benchmarks offer no guarantees of safety. You need a domain-specific assessment with expert validation. Without it, you won’t know exactly what clinically significant errors the watermark introduces into your text.

The next question to ask watermarking scheme developers is: Is it possible to create a watermark that preserves semantics in domains where each token is highly valuable? There is no answer yet.

Source: cs.AI updates on arXiv.org