Watermarking LLMs for Medical Use Backfires, 11-Model Study Reveals

Researchers from ETH Zurich and other institutions have published the first rigorous study evaluating the impact of LLM watermarks on medical performance. Watermarking is widely proposed to trace AI-generated content, but most evaluations rely on general-purpose benchmarks that ignore domain-specific risks. The team tested five watermarking schemes across 11 large language models and 7 vision-language models on a range of unimodal and multimodal clinical reasoning tasks. They introduced a human-expert-validated pipeline to systematically audit medical reasoning quality, terminological precision, and induced hallucinations.

The findings reveal that watermarking can cause substantial degradation in medical outputs, including lexical corruption, hallucinated terminology, and amplified misattribution or omission of image findings. Crucially, the study demonstrates that aggregate metrics like BLEU or perplexity systematically obscure these clinically consequential failures. The authors conclude that domain-specific evaluation is a prerequisite for safe deployment of watermarked models in medicine, as current benchmarks risk masking errors that could harm patients.