What Is textGrain? OpenAI's LLM Watermarking for AI Text: How It Works, Coverage and Accuracy (2026)

Will text written with ChatGPT become identifiable later as "written by AI"? On October 5, 2026, OpenAI announced textGrain, a technology that adds an invisible watermark to the text it generates. It is starting the rollout in ChatGPT and Codex, beginning with users in the EU (European Union). Based on OpenAI's official announcement and technical report, this guide covers how textGrain works, which services it covers, and its detection accuracy and limitations, as of October 9, 2026.

What textGrain is

textGrain is a technology that adds a statistical mark, imperceptible to people, to text generated by OpenAI's AI. A dedicated detector looks for that mark and determines whether the text contains OpenAI's watermark. Watermarks added to AI text like this are generally called LLM watermarking.

OpenAI announced the technology as part of its response to the EU AI Act. The AI Act requires providers of generative AI to make text and other content created by AI detectable by machines.

Note that Textgain (without an "r"), which has a similar name, is a Belgian AI company and has nothing to do with textGrain.

How it works: leaving a mark in word choice

Rather than inserting symbols or invisible characters into the text, textGrain leaves its mark in which words the AI chooses. According to the technical report, the mechanism works roughly as follows.

A random number is generated from a secret key and the preceding context
When the AI chooses the next word, the choice is adjusted so that words aligned with that random number become slightly more likely
The adjustment is designed not to change the original probabilities on average, preserving the quality of the text
The strength of the watermark is tuned by a "budget" for how much text diversity may be sacrificed

Detection requires only the text and the secret key. There is no need to run the original AI model.

OpenAI says that in its own evaluations, textGrain performed as well as or better than other methods, including Google's SynthID Text. It plans to release the technical details as open source in the future.

Impact on quality

OpenAI says it saw no meaningful difference in response quality with the watermark added. In the published benchmark results, the differences were small: for example, GPQA Diamond, a set of expert-level science questions, went from 94.44% to 93.94%, and BrowseComp, which measures browsing ability, went from 87.92% to 87.35%.

Coverage and rollout

The scope of textGrain differs by service.

ChatGPT and Codex: over the coming weeks, text for eligible users on all plans in the EU will be watermarked. Outside the EU, OpenAI says it currently has no plans to make it the default.
API: from October 5, 2026, API users worldwide can turn on the watermark for some models (off by default).
Via the cloud: OpenAI is preparing to make the watermark available within a few weeks for its models used through cloud providers as well.

The names of the covered models and the details of how to configure it in the API had not been published at the time of the announcement.

Who can detect the watermark

The detector that checks text for the watermark is not publicly available. At first, only vetted researchers and specialist organizations can apply, and access is decided case by case. OpenAI says the detector returns only "whether OpenAI's watermark is present" and does not reveal the user, the prompts entered or the content of the conversation.

For images and audio, on the other hand, anyone can continue to check whether they were created with OpenAI's AI using openai.com/verify or the Content Provenance API.

Detection accuracy

OpenAI has also published figures on watermark detection accuracy. With false positives (the rate at which human text is judged to be watermarked) held to 1%, the results were as follows.

Text length: for text in fields such as psychology, the watermark was detected in about 80% of texts at around 200 tokens and about 95% at around 400 tokens
Differences by field: for text with little freedom in word choice, such as mathematics, the detection rate dropped significantly
Rewriting: in 400-token texts, replacing 10% of the words with synonyms lowered the detection rate from about 92% to 66%, and replacing 25% lowered it to 17%

A token is the unit in which AI processes text. In English, one word is roughly one token.

What the watermark cannot tell you

OpenAI also states clearly what the watermark result cannot reveal.

The degree of human involvement: it can show that AI was involved, but not how much editing or creative work a person added
Ownership or responsibility: it cannot determine who owns the text or whether its use was lawful
The user: it does not show which person or account created it
Accuracy of the content: it cannot tell whether the text is accurate or misleading
No watermark does not mean a person wrote it: the watermark may not be detected because the text is too short, was edited or translated, was created with a model not covered, was created before the watermark was introduced, or was created with another company's tool

How it differs from SynthID Text

Another well-known watermark for AI text is Google's SynthID Text.

SynthID Text: released as open source by Google in 2024, so anyone can build it into their own model. It uses "tournament sampling," which picks words through a knockout competition
textGrain: a watermark OpenAI adds to the output of its own services. Access to the detector requires vetting, and open-sourcing is planned for the future

Both leave their mark in word choice. Both also share the same limitation: they are weak against short text, heavily rewritten text and translated text. For more, see "What is SynthID Text?", "What is LLM watermarking?" and "AI watermarks compared".

Impact on users in Japan

For now, the watermark in ChatGPT and Codex applies only to users in the EU. This announcement does not say that text written in ChatGPT by users in Japan will be watermarked. If you are developing with the API, you can turn on the watermark yourself.

Note that how well the watermark can be detected in Japanese text has not been published. OpenAI also acknowledges that translation weakens the watermark.

Summary

textGrain is a watermark for AI text that OpenAI introduced to comply with EU regulation. The rollout begins with ChatGPT and Codex in the EU, and it is available optionally worldwide in the API. Its impact on quality is small, but it is weak against short text and rewriting, and for the time being the detector is limited to vetted researchers and similar parties.

A watermark is only one clue as to whether text was created with AI. Do not judge whether a person wrote something based on the presence or absence of a watermark alone.