A Watermark for Black-Box Language Models

Dara Bahri, John Wieting

Watermarking has recently emerged as an effective strategy for detecting the outputs of large language models (LLMs). Most existing schemes require white-box access to the model's…