How Good is Post-Hoc Watermarking With Language Model Rephrasing?

Abstract

Text watermarking embeds statistical signals into text that can later be detected, enabling traceability of AI-generated content. We explore post-hoc text watermarking through LLM rephrasing, where watermarks are embedded during the rewriting process of existing text to, for instance, protect copyrighted documents or detect their use during training or RAG. Unlike generation-time watermarking which is constrained by how LLMs are served in practice, the post-hoc setting offers control over many generation and detection parameters. We test if we can exploit this freedom by allocating more compute to get a better quality-detectability trade-off, ranging from using bigger models for rephrasing, using beam search or even generating multiple candidates, to utilizing an auxiliary model for entropy filtering at detection time. Our results show that these strategies achieve strong detectability and high semantic fidelity on open-ended text like Wikipedia articles and books, while verifiable text like code remains more challenging due to stricter correctness constraints.