Researchers have developed a new approach to detecting statistical watermarks in language models. These watermarks are hidden within the model's output and can be used to verify its authenticity. However, they are vulnerable to certain types of transformations that preserve the meaning of the text but alter its form. The authors propose a new method for measuring these transformations, which takes into account not just the endpoint of the transformation but also the 'holonomy' or internal structure of the transformation itself. This approach is based on the concept of linguistic loops and allows for a more accurate detection of statistical watermarks. --- Why it matters: This research matters to AI engineers because it addresses a critical issue in the development of language models: the vulnerability of statistical watermarks to certain types of transformations. By providing a new method for measuring these transformations, the authors' work has implications for the security and integrity of AI-generated content. Source: https://arxiv.org/abs/2608.19369