What is OpenAI's textGrain watermark, and what does it change?

OpenAI published three things on October 5, 2026. A post titled "Our approach to EU text provenance rules" announced the watermark and its scope. Its help article on provenance signals was retitled and extended to cover text. And a technical report, "textGrain: Entropy-Calibrated Watermarking for Language Model Text," described the method, written by researchers from the University of Pennsylvania, Yale and OpenAI.

The post's central sentence: "Over the coming weeks, we will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union." It also says OpenAI is "not making text watermarking a global default at launch."

This is OpenAI's first shipped text watermark. Reporting in 2024 described an earlier internal system, said to be about 99.9% effective, that OpenAI built and chose not to release over false positives, circumvention, the effect on non-native English writers, and a survey suggesting about 30% of users would use ChatGPT less. Until 2026-10-05 OpenAI marked images and audio but no text.

API customers turn it on under Organization settings > Data controls > Text provenance: tick "Allow text watermarking," pick models, and save. Projects can override the organization setting under Project Settings > Text provenance. There is no API parameter for it.

OpenAI has not published a list of eligible models; it is shown in Settings. It says it will extend coverage to all legacy models over the coming weeks and that the speed impact is negligible. It gives no pricing information, and turning watermarking on does not give a customer access to the detector.

A language model writes one token at a time. At each step it has a probability for every possible next token, and it normally picks one at random according to those probabilities. textGrain replaces that ordinary randomness with randomness derived from a secret key and the text just before the current position.

The amount of steering is set as an entropy budget. Entropy here measures how uncertain the model is about the next token. textGrain is allowed to remove a fixed fraction of that uncertainty and no more. When the model has many reasonable choices, there is room to steer; when it is nearly certain, as in a calculation or exact code, there is little room, which is why OpenAI reports weaker detection on mathematics and code.

The report describes the steps in more technical terms.

• Token sampling is coupled to the keyed randomness by solving an optimal-transport problem with Gumbel costs and a KL-divergence penalty. The KL term equals the entropy removed, which is what makes the budget exact.
• The vocabulary is split into blocks using the key. The transport problem is solved on block probabilities with a Sinkhorn solver, a block is chosen, and tokens inside that block keep their original relative probabilities.
• The scheme is unbiased in the technical sense used in watermarking papers, and it is compatible with speculative sampling, a common way of speeding up generation.
• Detection needs only the text and the key, not the model. It scores the first occurrence of each distinct context window, and under the assumption of unwatermarked text the score follows a Gamma(n, 1) distribution, so a threshold gives a chosen false positive rate under idealised independence. The report says that rate needs empirical calibration in practice.

The report contains no empirical results. Every detection and robustness number on this page comes from OpenAI's post and help page, and OpenAI says the report will be updated with additional details in the coming weeks.

OpenAI's figures are measured at a 1% target false positive rate, meaning the threshold is set so that about 1 in 100 unwatermarked passages would be wrongly flagged. On English ELI5 data, about 80% of 200-token passages and about 95% of 400-token passages were detected, for content such as psychology. OpenAI says mathematics is substantially lower and gives no number.

Other EU languages score lower. At the same 1% rate, Spanish was highest at 69.0% and Romanian lowest at 42.2%. OpenAI says it raised the watermark strength for languages that fell below 60%.

OpenAI's help page lists the weak cases: short passages are unreliable, code is harder to watermark, and factual or precise answers and verbatim reproduction carry a weak signal.

On output quality, OpenAI compared its latest frontier model, Astra, with and without the watermark: Terminal-Bench Science 56.90% watermarked against 60.00% without, DeepSWE 72.80% against 71.68%, and GPQA 94.44% against 93.94%. OpenAI reads the differences as noise.

OpenAI tested synonym replacement on 400-token passages. Replacing 10% of the words with synonyms cut detection from about 92% to 66%. Replacing 25% cut it to 17%.

The help page says substantial paraphrasing or translation can make the watermark undetectable, and that it survives light edits, copy and paste, and screenshots, depending on the content. Short text is unreliable to begin with.

Copy and paste does not affect it because there is nothing to strip. OpenAI says textGrain inserts no hidden characters or invisible spaces, so character cleaners have no effect on it either way.

OpenAI's post lists what a detection does not do.

• It does not measure how much a human contributed.
• It does not establish ownership or responsibility.
• It does not identify the user. The detector reports whether it detects an OpenAI watermark "without identifying the user or revealing their prompts or conversations."
• It does not verify accuracy.
• It cannot tell whether the model wrote the text or only edited text someone uploaded. The help page calls a watermark evidence that an OpenAI model "likely generated or processed" the content, and says it does not replace visible AI labels.

The post also says: "The absence of a detected watermark does not prove human authorship." The text may be too short, edited, translated, produced by an unsupported model, written before watermarking began, or produced by another company's tool.

The detector itself is limited at launch to approved researchers and expert organizations, case by case, through an application form. OpenAI named John Thickstun (Cornell), Martin Vechev (ETH Zurich and INSAIT) and researchers at KInIT as partners. If you are not approved, you cannot test text yourself. OpenAI says it plans to open-source textGrain.

Google documents SynthID Text as a logits processor applied during generation and has open-sourced a version with a reference detector. OpenAI says textGrain "matched or exceeded" other approaches including SynthID for text in its tests, and that it was built for more control over the trade-off between detectability and response diversity than SynthID or TextSeal. Those are OpenAI's own comparisons; no independent comparison has been published.

Anthropic describes Claude's watermark as a version of the SynthID-Text approach, keyed word selection with nothing added to the text, applied at the model level, so text from a supported model carries it wherever Claude is offered. Its detector shipped on September 1, 2026 in private preview for eligible organizations.

The practical difference for a reader is geography. Text from a supported Claude model is marked wherever it was written. ChatGPT text is marked only if it was written in the EU, or through an API account that opted in.

OpenAI's post says the EU AI Act "requires generative AI providers to make generated text identifiable in a machine-readable way." Its help page cites the EU Code of Practice on Transparency of AI-Generated Content, which OpenAI signed. Neither names Article 50 or a date. On this site's reading, the relevant rule is Article 50(2), which applied from August 2, 2026, so textGrain arrived about two months after it.

OpenAI's help page notes that the Code of Practice does not require watermarks for outputs under 200 tokens, about 150 English words, or for code snippets. That lines up with where textGrain is weakest.

Limiting the watermark to the EU follows from the legal reason OpenAI gives. It also means the rest of the world's ChatGPT text carries no OpenAI text watermark at launch.