[2301.10226] A Watermark for Large Language Models

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 whitelist tokens before a word is generated, and then softly promoting use of whitelist 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.

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 whitelist tokens before a word is generated, and then softly promoting

related reading

• Scalable watermarking for identifying large language model outputs | Naturenature.com
• Large Language Model: world models or surface statistics?thegradient.pub
• Language Modelinglena-voita.github.io
• GitHub - salesforce/AuditNLG: AuditNLG: Auditing Generative AI Language Modeling for Trustworthiness · GitHubgithub.com
• Quantifying Truesight With SAEs · Gwern.netgwern.net
• [2310.18512] Preventing Language Models From Hiding Their Reasoningar5iv.labs.arxiv.org
• Red Teaming Language Models with Language Models — Google DeepMinddeepmind.com
• Making Large Language Models work for yousimonwillison.net
• 2506.17298arxiv.org
• How ChatGPT Could Embed a ‘Watermark’ in the Text It Generates - The New York Timesnytimes.com
• [2503.10965] Auditing language models for hidden objectivesarxiv.org
• Things we learned about LLMs in 2024simonwillison.net