A Watermark for Large Language Models
This addresses the issue of identifying AI-generated content for users concerned with misinformation or misuse, though it is incremental as it builds on existing watermarking concepts.
The paper tackles the problem of mitigating potential harms from large language models by proposing a watermarking framework that embeds detectable signals into generated text with negligible impact on quality, and demonstrates its effectiveness using a multi-billion parameter OPT model.
Potential harms of large language models can be mitigated by watermarking model output, i.e., embedding signals into generated text that are invisible to humans but algorithmically detectable from a short span of tokens. We propose a watermarking framework for proprietary language models. The watermark can be embedded with negligible impact on text quality, and can be detected using an efficient open-source algorithm without access to the language model API or parameters. The watermark works by selecting a randomized set of "green" tokens before a word is generated, and then softly promoting use of green tokens during sampling. We propose a statistical test for detecting the watermark with interpretable p-values, and derive an information-theoretic framework for analyzing the sensitivity of the watermark. We test the watermark using a multi-billion parameter model from the Open Pretrained Transformer (OPT) family, and discuss robustness and security.