Watermarking AI-generated text to help identify its origins significantly degrades the quality of the output, according to a new study. The research, which evaluated several watermarking techniques, found that the more robust the watermark, the worse the text quality becomes.
Trade-off Between Detection and Quality
The study, conducted by researchers at the University of Maryland, tested various watermarking methods on outputs from large language models. It found that watermarking can reduce the quality of the text by as much as 20%, with the most effective detection methods causing the greatest degradation.
Watermarking works by subtly altering the statistical patterns of the generated text, making it possible to trace its AI origin. However, these alterations can make the text less coherent, less fluent, and less accurate, the researchers said.
Implications for AI Regulation
The findings have significant implications for policymakers and tech companies, who have been exploring watermarking as a way to combat the misuse of AI-generated content, such as disinformation and academic cheating. The study suggests that any such measures must balance the need for detection against the potential harm to the quality of AI outputs.
The researchers also noted that some watermarking techniques could be easily removed by paraphrasing, limiting their effectiveness in real-world scenarios.
Call for Further Research
The study's authors call for more research into watermarking methods that minimize quality degradation while maintaining robust detection. They also suggest that watermarking may be more suitable for certain types of content, such as short snippets or code, where the impact on quality is less pronounced.
