On October 6, 2026, Beijing time, OpenAI announced its textGrain text-watermarking technology and a phased deployment plan. Starting October 5, API customers worldwide can opt in to watermarking for select models, while the API default remains off. Eligible ChatGPT and Codex text output in the EU will receive invisible watermarks over the coming weeks across all plans. What is in effect today is the API opt-in and the detector application process; the EU product rollout is not yet complete, and OpenAI has not made text watermarking a global default.
API opt-in starts first, EU products are covered later
OpenAI divides the product arrangement into three parts. API customers can choose to allow text watermarks and specify models in project or organization settings, but the official announcement does not list the first batch of supported models; in the EU, eligible ChatGPT and Codex users will receive watermarks in the following weeks; text detectors only accept applications from researchers and expert institutions, and are approved on a case-by-case basis. Turning on API watermarking does not automatically grant detector permissions, and public verification tools for images and audio do not extend to text.
OpenAI linked the deployment to the EU Artificial Intelligence Act requiring generative AI providers to make generated text legible by machines, while saying text watermarking and detection are still in their early stages. The detector may determine that text without a watermark is watermarked, or it may miss the signal that has been added. Short texts, content with restricted wording, areas with smaller word choices such as mathematics, and subsequent rewriting or translation will all reduce detection capabilities, so the company did not make the text detector available to the public at launch.
Watermark changes wording choices instead of inserting hidden characters
textGrain does not add hidden characters, invisible spaces, unusual punctuation, or extra watermark-only tokens to the output. OpenAI's help documentation says it slightly adjusts the model's random choice among possible words or word pieces, creating a statistical pattern that a detector with the corresponding key can look for. The watermark is part of the wording itself: copying and pasting does not introduce hidden content, while substantial rewriting, paraphrasing, or translation can make the signal undetectable.
A technical report released on October 5 describes textGrain as an entropy-calibrated scheme. The system uses a key and preceding context to generate pseudorandom values, divides the vocabulary into blocks, and arranges block selection through optimal transport with KL regularization, setting a budget between watermark strength and reduced sampling randomness. The detector reconstructs statistical scores from the same key, tokenizer settings, and context rules. The report also states assumptions about independence, keyed randomness, numerical solving, and calibration; its theoretical derivation is not an absolute guarantee for real-world text.
Length, language and paraphrase change detection performance
OpenAI's English-language evaluation reports that, at a target false-positive rate of 1%, the detector identified watermarks in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology. In another 400-token evaluation, detection began at about 92%; replacing 10% of the words with synonyms reduced it to 66%, and replacing 25% reduced it to 17%. These are OpenAI's self-reported results and were not independently reproduced in the materials reviewed here. The company also says detection was substantially lower for constrained content such as mathematics.
OpenAI's help documentation reports a language evaluation covering all 24 official EU languages. The test began with 500 synthetic English prompts and translated them into the other 23 languages. At a 1% target false-positive rate, Spanish had a 69.0% detection rate and Romanian 42.2%; OpenAI says watermark strength can be adjusted for languages below 60%. Because the test used specific prompts, translations, and settings, its results cannot be generalized to every language, model, domain, or text length.
Testing conclusion boundaries and undisclosed matters
As of this review, OpenAI had not published the initial API model list, country-by-country EU dates, detailed eligibility rules, the names of cloud or distribution partners, or a date for public detector access. Applying for access does not guarantee approval, and the reviewed materials do not include an external independent replication. Editorial analysis: textGrain places a machine-readable provenance signal in the generation process, but OpenAI's own data shows that length, domain, language, and rewriting change detection results, so the signal cannot replace fact-checking, authorship analysis, or legal responsibility determinations.