User-Controlled Post-Hoc Watermarking for LLM-Generated Text Without Model Access

Current watermarking approaches for LLM-generated text typically embed watermarks during the generation process, using a secret key controlled by the LLM. Although this enables verification of LLM-generated content, it may raise copyright and authorship concerns. For instance, when a user relies on an LLM for proofreading, grammar correction, or translation, the model may still embed its own watermark. This can later create disputes regarding ownership, as the text may appear to belong to the AI system rather than the human author. To address this issue, this paper proposes an improved post-hoc watermarking approach that embeds user-controlled watermarks after the text generation is complete. The approach does not require access to the internal architecture of the LLM. It applies a probabilistic watermarking mechanism using a user secret key, and embeds watermark signals across the entire text rather than only on selective terms, as in previous post-hoc approaches. During watermark embedding, the approach regenerates the suffix after each token substitution by interacting with the LLM. This preserves grammatical correctness and contextual coherence, addressing the quality degradation often observed in synonym-based substitution approaches that do not adjust the surrounding text. A BERT-based masked language model is used to ensure that each substitution remains semantically consistent with the original text. Experimental results show that the proposed approach maintains higher similarity and readability than previous post-hoc approaches while producing statistically verifiable watermark signals. The approach provides a practical solution for authorship verification and copyright protection in generative AI, without requiring access to the internal mechanism of the LLM.