A recent arXiv paper, "The Feasibility of Topic-Based Watermarking on Academic Peer Reviews," takes up the question of how to detect large language model (LLM) involvement in peer review. The abstract notes that LLMs are now commonly permitted by conferences and journals for tasks like language refinement and literature summarization, which raises the question of their use in writing reviews themselves.
The paper appears to propose a watermarking technique based on topic, rather than on text statistics or other common markers, and evaluates whether such an approach is feasible for academic peer reviews. The abstract is truncated, so specific methods and results are not available from the source alone.
What is clear is that the paper frames this as a practical concern: as LLM use becomes routine in academic workflows, tools to trace AI-generated review text are needed. The study's focus on topic-based watermarking suggests an attempt to make detection robust even when the text is heavily edited or paraphrased, though the source does not confirm those details.
Overall, the paper contributes to the growing discussion around AI accountability in scholarly publishing, but the full findings will require reading the complete manuscript.