A new arXiv paper examines a practical risk in watermarking AI-generated medical text: the mark itself may change meaning. Watermarks are designed to make model output traceable, often by nudging token choices in ways that remain hard for readers to notice.
The authors argue that medicine deserves special testing because small wording changes can matter. They benchmark five watermarking schemes across 11 large language models and seven vision-language models on medical tasks.
The study is important because hospitals, insurers, and health software vendors may want traceability as AI enters clinical workflows. But a watermark that slightly reduces accuracy or shifts terminology could be unacceptable in settings where a detail affects diagnosis, treatment, or patient instructions.
The paper does not mean watermarking should be abandoned. It shows that domain-specific evaluation is necessary. A method that looks safe on general writing benchmarks may behave differently when the text contains symptoms, dosages, or medical reasoning that must remain precise.