OpenAI Launches textGrain Watermarking to Mark AI-Generated Text
OpenAI has officially released textGrain, an entropy-calibrated text watermarking technology, alongside a technical paper published on its blog. The method embeds invisible statistical signals during model text generation without altering the wording or punctuation, allowing machines to identify whether text was produced by OpenAI models.
According to OpenAI's deployment plan, the watermark will be enabled by default in ChatGPT and Codex within the European Union, while global API access will leave the choice to enterprises. The watermark detector is not open to the general public and is currently available only to approved researchers and institutions. Su Weijie, the paper's corresponding author and OpenAI researcher, said on X that the textGrain algorithm will later be open-sourced for the global research community to improve.
Existing watermarking approaches typically hide visible symbols or special characters, but textGrain instead guides the probabilistic sampling process during generation. A key generated from context creates pseudo-random rules that subtly influence word selection, producing statistical patterns detectable by algorithms but imperceptible to human readers. Li Xiang, first author of the paper and a postdoctoral researcher at the University of Pennsylvania, explained that the method only modifies sampling logic and is independent of model pretraining, making it theoretically adaptable to multiple languages. However, he noted that parameters tuned for English would require recalibration for Chinese due to differences in vocabulary, conventions and statistical features.
OpenAI said internal tests showed no practically significant performance degradation on the frontier model Astra when the watermark was enabled across multiple benchmark suites. The company also acknowledged clear boundaries for the technique: ideal-lab performance does not guarantee reliable detection in everyday use. Rewriting and synonym substitution can substantially reduce detection rates, and the watermark cannot determine the proportion of AI-generated versus human-written content, establish copyright or legal liability, track specific accounts, or verify factual accuracy. The tool is intended as a traceability clue rather than a definitive judge of authorship.
