ChatGPT watermark: how textGrain works

On October 5, 2026, OpenAI announced that it would add a watermark to ChatGPT and Codex text produced in the European Union. This watermark is called textGrain and it is invisible. There is no hidden character and no special space: the mark lies in the choice of words, and only OpenAI's detector can read it.

The rollout is happening "in the coming weeks", across all plans. Outside Europe, nothing changes for ChatGPT users. On the API side, watermarking is available worldwide, but it is turned off by default.

This change of heart is anything but spontaneous. OpenAI had a watermark ready as early as 2023 and kept it on the shelf for three years. It is bringing it out because the AI Act requires it: its transparency rules have applied since August 2, 2026.

What OpenAI announced on October 5, 2026

Only ChatGPT and Codex text "generated in the EU" will be marked (OpenAI, TechCrunch, BleepingComputer). OpenAI does not say how it sorts this out: account country, connection location, something else? Nobody knows.

On the API, watermarking can be turned on since October 5 for some models, in the project or organisation settings. If the developer does nothing, nothing is marked, whether they call the API from Paris or from San Francisco.

The detector itself is not public. For now, OpenAI reserves it for researchers and organisations it approves case by case. OpenAI also warns that the absence of a watermark does not prove that a human wrote the text, and that the mark does not measure the share of human editing or creativity.

What is textGrain and how does it work?

textGrain is the name of OpenAI's text watermark, presented on October 5, 2026 in a technical report co-authored with the University of Pennsylvania and Yale. It is a statistical mark, invisible when reading, embedded in the choice of words made by ChatGPT and Codex, which only OpenAI's detector can read.

A language model writes token by token (a token is a word or part of a word, see our explanation of the key concepts of AI). At each step, it computes a list of possible continuations with their probabilities, then draws one at random.

The watermark acts on this draw. The model does not change what it means to say: it replaces chance with pseudo-chance, computed from a secret key and the few previous tokens. To a reader, nothing distinguishes the result from a normal draw.

Whoever holds the key can redo the calculation. They check, word after word, whether the text's choices follow the expected pseudo-random sequence. A single choice proves nothing. Hundreds of choices all pointing the same way do. It is a statistical test, whose false positive rate is set in advance: on human text, the correlation is zero on average (textGrain technical report, OpenAI, University of Pennsylvania and Yale).

The principle is not new. Scott Aaronson, then at OpenAI, proposed it in 2022. A team at the University of Maryland formalised it in 2023 (Kirchenbauer et al.), and Google DeepMind has used it in Gemini since 2024 under the name SynthID-Text (Nature). textGrain adds an adjustable "entropy budget", which sets how much mark is embedded depending on how much freedom the model has.

And answer quality? Averaged over all possible keys, the word distribution does not change. According to figures reported by AI Weekly, the benchmark score goes from 49.76 without the mark to 49.57 with it. Google had already tested SynthID-Text on nearly 20 million Gemini answers, with no difference in satisfaction between marked and unmarked answers.

No, the ChatGPT watermark is not an invisible character

Search for "ChatGPT watermark" on Google: most results tell you that ChatGPT slips invisible spaces into its text, and offer a tool to remove them. This is wrong, or at least it has nothing to do with textGrain.

OpenAI says so explicitly: its provenance signals use neither invisible characters nor special punctuation (OpenAI help page). A watermark made of hidden characters would be useless anyway: the first find-and-replace would erase it.

The non-breaking spaces sometimes found in ChatGPT's answers are typography. In French, they are even the rule: a non-breaking space goes before colons, semicolons, exclamation marks and question marks. A well-typeset human text in French contains dozens of them. As for the em dash, humans used it long before language models.

The "watermark removers" that clean up these characters therefore miss the point: the watermark is in the words themselves.

Who is affected?

You use ChatGPT from Europe (on chatgpt.com or in the app): your text will be marked once the rollout is complete, whatever your plan. There is no setting to turn it off.

You use ChatGPT outside Europe: nothing changes.

You build with the OpenAI API: you decide. Watermarking is off by default, worldwide, and can be turned on in the project or organisation settings. This is the big difference with chatgpt.com, where European users have no choice.

You use Codex: same rule as ChatGPT, code and text generated in the EU will be marked. In practice, code that has to be exact leaves the model little freedom, and so little room for the mark.

You use Copilot: Microsoft relies heavily on OpenAI models, but nothing public indicates so far that Copilot text inherits the mark.

Can you detect text written with ChatGPT?

With the watermark, yes, provided you have access to OpenAI's textGrain detector and a text that is long enough and lightly edited. The figures published by OpenAI and quoted by BleepingComputer:

Some texts lend themselves poorly to watermarking. A short text does not contain enough choices to build up a signal, and the European code of practice does not require a watermark below 200 tokens. An answer with no room for manoeuvre (a fact, a calculation, a minor correction) offers nothing to mark, since only one wording is correct. Finally, a translation changes every word.

OpenAI's watermark only sees OpenAI's text. It says nothing about a text written with Claude, Gemini, Mistral or an open-weight model that someone runs on their own servers.

What about classic AI detectors?

Pangram, GPTZero, Turnitin or ZeroGPT work differently: they do not have the key, they guess from style. The watermark does not help them, since it does not change the style they analyse.

Their reliability varies a lot. OpenAI's own detector, withdrawn in July 2023, caught only 26% of AI texts and wrongly flagged 9% of human texts. The best commercial tools do much better on test corpora (University of Chicago, 2025), but no score proves anything on its own. The exclusion of Thélyson Orélien's novel from the Goncourt Prize selection, in September 2026, was a reminder: only Pangram concluded the text was AI-written, other detectors judged it human, and the Académie also relied on the author's old, proven plagiarism (ActuaLitté).

Can you remove the ChatGPT watermark?

Yes, and without much effort. The watermark is there to trace good-faith use; it does not withstand someone who wants to make it disappear.

The simplest way is to go through the OpenAI API. On chatgpt.com, European users have no choice; on the API, the developer decides whether to turn watermarking on, and it is off by default. Text generated through the API without this option carries no mark, even from Europe.

The other methods go through the words:

• rephrase: replacing a quarter of the words drops detection to 17%, and a full rewrite, by hand or with another model, erases it;
• translate: a round trip through another language changes every word;
• switch models: open-weight models served by a third party, or that a company runs on its own servers, mark nothing by default.

The AI Act places obligations on providers and on companies that deploy AI systems, not on individuals: it does not forbid a user from removing a mark, unlike Chinese law. On the other hand, a company that integrates the OpenAI API into a service offered in Europe remains bound, as a provider, to mark the text it generates. For that company, OpenAI's option is the simplest way to comply.

And removing a watermark changes nothing about the rules a school, a publisher or an employer sets on the use of AI.

ChatGPT, Claude, Gemini, Mistral: who marks what?

Sources: Anthropic, Claude help, Google DeepMind, Microsoft, European Commission.

Between Claude and ChatGPT, the gap is clear. Anthropic marks everything, everywhere, for lack of a reliable way to limit marking to one region. OpenAI marks the minimum the law requires: ChatGPT in the EU, the API only if the developer asks for it.

Why now: the AI Act

Article 50 of the AI Act has applied since August 2, 2026. Its paragraph 2 requires providers of AI systems that generate content, text included, to mark their outputs in a machine-readable and detectable format. Solutions must be "effective, interoperable, robust and reliable" as far as technically feasible. Standard editing functions (spell checking, basic translation, formatting) are exempt.

The timeline has been adjusted: systems already on the market before August 2 have until December 2, 2026 to mark (CSA). Penalties can reach €15 million or 3% of worldwide turnover.

A code of practice, finalised in June 2026, sets out the expectations: a watermark for texts longer than 200 tokens, a free detector at least for regulators, researchers and civil society. Around 190 organisations have signed it, including OpenAI, Anthropic, Google, Microsoft and Mistral (European Commission).

Along with China, the European Union is the only jurisdiction requiring text to be marked. Elsewhere, laws primarily target deepfakes: image, voice, video.

The watermark OpenAI kept on the shelf

In August 2024, the Wall Street Journal revealed that OpenAI had had a text watermark for about a year, 99.9% effective according to internal documents, without deploying it. An internal survey showed that nearly 30% of users would use ChatGPT less if its text were marked. OpenAI confirmed it, also citing how easy it is to get around and the risk of stigmatising non-native English speakers (Search Engine Journal).

Two years later, the watermark is arriving, but only where the law requires it.

Frequently asked questions about the ChatGPT watermark

Does ChatGPT watermark its text?

Yes, in the European Union. Since OpenAI's announcement on October 5, 2026, ChatGPT and Codex text generated in the EU is gradually receiving an invisible watermark called textGrain, on all plans. The watermark is statistical: it slightly steers the choice of words, without adding any hidden character. Outside Europe, ChatGPT does not mark its text.

Is the watermark turned on in the OpenAI API?

No, not by default. On the API, the developer chooses whether to turn the watermark on, in their project or organisation settings, for some models. This is the difference with chatgpt.com, where European users cannot turn it off. A company offering a service built on the API in Europe is nevertheless bound by the AI Act to mark the text it generates.

Can you remove the watermark from ChatGPT text?

Yes. The watermark disappears when the words change: replacing a quarter of the vocabulary drops detection to 17% according to OpenAI, and a full rewrite or a translation erases it. Text generated through the API without turning the option on is not marked. Tools that "clean" invisible characters are useless against textGrain, which does not rely on those characters.

Are the invisible spaces in ChatGPT text a watermark?

No. OpenAI states that its provenance signals use neither invisible characters nor special punctuation. Non-breaking spaces are a matter of typography, and French requires them before colons, semicolons, exclamation marks and question marks. Their presence proves nothing about where a text came from.

Can a teacher tell that a text was written with ChatGPT?

Not with the watermark, for now: OpenAI's detector is reserved for approved researchers and organisations. AI detectors available to the public analyse style and give a probability, not proof. The most solid evidence remains the trail of the work: drafts, successive versions, notes.

Does the watermark reveal my identity or my account?

No. According to OpenAI, the watermark reveals neither the author of the text, nor their account, nor the prompt, nor the conversation. It only indicates that an OpenAI model probably generated all or part of the text, without being able to say which part was written or edited by a human.

Do Claude and Gemini also watermark their text?

Yes. Gemini has used SynthID-Text since 2024. Claude has marked all its text since August 2026, worldwide and with no way to turn it off, using a method derived from SynthID-Text. Each watermark can only be read by its own provider's detector: Google's detector does not see Claude's mark, and vice versa.

---

In Europe, a long text copied as is from chatgpt.com becomes identifiable by those with access to the detector. That is about as far as it goes. The watermark disappears with rephrasing, covers the API only if the developer wants it to, ignores other models and proves nothing when it is missing.

For now, OpenAI ticks the box Brussels requires, and no more. The system will only really be useful once detectors are open and able to read the marks of several providers, which the European code aims for by 2027. Until then, a detector score remains a clue, and dated drafts are still worth more than a percentage.