OpenAI rolls out text watermarking system textGrain for AI-generated writing

OpenAI has introduced textGrain, a text watermarking system designed to help identify whether a passage was generated by its models without making the writing look obviously altered. According to the report cited in the source article, the feature is available immediately for global API customers, while ChatGPT and Codex are set to adopt it in the coming weeks. OpenAI said that with the false positive rate held at 1%, more than 90% of 400-token psychology-style responses can be detected. The company also said performance impact is minimal: on eight core benchmarks for GPT-6 Astra, results were broadly unchanged and five scores improved slightly. The article traces textGrain’s technical roots to optimal transport, a mathematical framework dating back to Gaspard Monge’s 1781 earth-moving problem and later expanded by Leonid Kantorovich. It says the method allocates a “randomness budget,” allowing the model to create a detectable statistical link between token choices and a secret key while preserving the original token distribution on average. Detection, the article says, requires only the text and the key, not the model itself. The source also highlights the research team behind the 20-page report, including authors from OpenAI, the University of Pennsylvania, and Yale. Several key contributors, including first author Xiang Li and corresponding author Weijie Su, have academic ties to Peking University’s School of Mathematical Sciences.