Drs. Xiang Li, Qi Long, and Weijie Su have developed a novel approach for robust detection of AI-generated text in their paper entitled “Robust detection of watermarks for large language models under human edits,” published in the Journal of the Royal Statistical Society Series B: Statistical Methodology.
As generative AI, such as large language models (LLMs), transforms many sectors of our society—including medicine—detecting AI-generated text is becoming increasingly important. However, current state-of-the-art detection approaches often assume that a text is either entirely AI-generated or entirely human-written. In reality, many texts are mixtures of both, with humans partially modifying AI outputs. This study introduces a new, principled framework that models this real-world scenario as a mixture distribution and connects it with a classic idea in statistics—the goodness-of-fit test. This connection leads to detection methods that are both theoretically sound and practically reliable. The proposed tests achieve optimal robustness under two key criteria, ensuring efficient and robust detection even when texts are heavily or lightly edited.
Beyond these results, the work showcases how deep statistical insight can drive progress in AI—offering a rare combination of theoretical rigor and practical innovation to advance the reliable detection of AI-generated content.
This research was led by Dr. Xiang Li, who is a postdoctoral researcher, under the joint supervision of Dr. Weijie Su and Dr. Qi Long.