ChatGPT Watermark Is Official: Why Only the EU Gets It

Does ChatGPT have a watermark?

ChatGPT is getting a text watermark, and OpenAI confirmed it on October 5, 2026. The company says it will introduce watermarking for eligible ChatGPT and Codex users in the EU over the coming weeks, while API customers worldwide can already opt in. Outside the EU, ChatGPT text is not watermarked by default.

I published this post in May with "no" in the headline. That answer held for years and expired in one afternoon, when OpenAI posted "Our approach to EU text provenance rules". The key sentence, in OpenAI's words: "Over the coming weeks, we will introduce text watermarking to eligible ChatGPT and Codex users across all plans in the EU only. We are not making text watermarking a global default at launch."

So two popular claims are both wrong today. "ChatGPT never watermarks" is out of date, and "ChatGPT now watermarks everyone" was never true.

What is textGrain and how does it work?

textGrain is OpenAI's name for its text watermark, a statistical signal placed in the words the model picks as it writes. A secret key subtly shapes those word choices, and a detector holding the same key can measure the pattern later. Nothing is hidden in the file, the formatting or any invisible character.

OpenAI describes it this way: "Our text watermarking technology, textGrain, adds an invisible statistical signal to the model's word choices. Our detector looks for that signal to assess whether a passage contains an OpenAI watermark."

The idea belongs to a research family that became widely known with a 2023 paper by Kirchenbauer and colleagues, whose scheme selects a keyed set of "green" tokens and softly promotes them during sampling. textGrain is its own design, set out in OpenAI's technical report, written with university co-authors.

Because the signal is made of visible words, TechCrunch drew the obvious inference in its report: "Because it lives in the words themselves, it travels with the text when it's copied and pasted." That sentence is the reporter's reading, not a statement from OpenAI.

Length matters a great deal. Here are the numbers OpenAI published from its own evaluation (a token is a word or a piece of one):

OpenAI adds that detection rates were "substantially lower" for content such as mathematics. These are the vendor's figures on English text, and no outside lab has replicated them.

Who gets the ChatGPT watermark: the EU, API users or everyone?

The ChatGPT watermark goes first to eligible ChatGPT and Codex users in the EU, on all plans. API customers anywhere in the world can opt in for select models, and it is off unless they turn it on. Everyone else using ChatGPT outside the EU gets no text watermark by default.

Three groups, three different answers:

• ChatGPT and Codex in the EU. Watermarking arrives "over the coming weeks." OpenAI gives no start date and does not define "eligible."
• API customers worldwide. "Starting today, API customers around the world will be able to opt in to watermarked text outputs for select models," the October 5 post says. OpenAI frames this as letting customers decide how watermarking fits their own transparency obligations.
• ChatGPT outside the EU. No default watermark. OpenAI wrote that it is not making this a global default at launch.

The practical upshot: a paragraph can carry textGrain or not depending on where the account sits and whether a developer flipped a switch, which is why the text alone cannot tell you what produced it. There is no setting in ChatGPT to look up either, since OpenAI has not yet said which accounts count as eligible.

How do you check for a ChatGPT watermark? Is there a detector?

A ChatGPT watermark cannot be checked by the public today, because OpenAI has not released its detector. Access goes case by case to approved researchers and expert organizations. Reading the signal takes OpenAI's secret key, so a website that promises a textGrain scan has nothing to scan with.

OpenAI's stated reason is blunt: "Given the risk of missed watermarks and false positives, we are not making it publicly available at launch." For those who do get access, the tool "will report whether it detects an OpenAI watermark, without identifying the user or revealing their prompts or conversations." OpenAI's public verification page covers images and audio, not text.

What you can check for free is a different question: how your writing reads. Our free AI detector scores writing style and returns a score, a verdict and the patterns behind it, with 4 free scans a day and no account. It measures style, not provenance, and that is the reading most people actually face.

Is there a ChatGPT watermark remover?

A ChatGPT watermark remover that strips invisible Unicode characters cleans up a 2025 formatting quirk, and it does nothing to textGrain, which is made of ordinary visible words. There are no hidden characters in textGrain for such a tool to delete.

The invisible-character story dates to April 2025, when a startup called Rumi reported narrow no-break spaces (Unicode U+202F, or NNBSP) in text from OpenAI's o3 and o4-mini models and described them as a hidden watermark. OpenAI told Rumi, as later reported, that the characters were "simply a quirk of large-scale reinforcement learning." Narrow no-break spaces are also ordinary typography, used to keep words from splitting across lines.

Stripping stray characters is harmless housekeeping. Open the text in an editor that shows hidden characters, search for the narrow no-break space (U+202F) and replace it with a normal space. Just know what that buys you: cleaner formatting, with the wording, and any statistical signal in it, left exactly as it was.

One scope note, because the same search phrase is used for pictures: this post is about text. ChatGPT images carry Content Credentials and SynthID marks, which are a separate system with their own public checker at OpenAI's verify page.

How do you remove a ChatGPT watermark from text?

No method is proven to remove the ChatGPT watermark from text, because nobody outside OpenAI can test for it. The evidence points one way, though: the signal lives in word choices and weakens as those words change. OpenAI measured detection falling from about 92% to 17% when 25% of words were swapped for synonyms.

Research on older watermark schemes points both ways, and none of it tested textGrain:

• Sadasivan and colleagues found that recursive paraphrasing "can significantly reduce detection rates" for watermark detectors.
• Krishna and colleagues built a paraphrasing model whose output several detectors, watermarking included, failed to catch.
• A second Kirchenbauer paper pushed back: watermarks "remain detectable even after human and machine paraphrasing" once enough text is observed, on the order of 800 tokens after strong human paraphrasing.

So the accurate statement is that heavy rewriting weakens the signal, and the result depends on how long the text is and how much of it changes. Fixing a typo changes almost nothing.

Here is where our tool sits. Our humanizer does not swap a few synonyms. It rewrites your draft with our own model, trained on 2,590 real student essays, so the wording on the page is new rather than ChatGPT's. OpenAI's synonym test changed a quarter of the words and left the sentences in place; a full rewrite rebuilds the sentences themselves. No public textGrain detector exists, so no outside party can publish a textGrain score yet, and you should distrust any page that prints one.

Use it on your own drafts, in settings where AI help is allowed.

Do AI detectors read the ChatGPT watermark?

AI detectors do not read the ChatGPT watermark. OpenAI's report says its detector "requires only the generated text and the secret key," and OpenAI holds that key. The detectors used in classrooms and newsrooms score writing style instead, so a flag from one of them is a judgment about how the text reads.

I build the style model behind our own detector, and it has no channel for a keyed signal. It looks at how predictable the phrasing is, how the sentence rhythm moves and which patterns repeat. Our guide to what separates human text from AI text walks through those signals, and our page on how accurate AI detectors are covers how often style scoring gets it wrong.

The two systems answer separate questions. A watermark asks where the words came from. A style detector asks what the words look like. Text can pass one and fail the other in either direction, which is why a style flag is fixable by changing how the writing reads.

Does no ChatGPT watermark mean a human wrote it?

No ChatGPT watermark does not mean a human wrote the text, and OpenAI says so itself: "The absence of a detected watermark does not prove human authorship." The passage may be too short, edited or translated, or it may come from an account or a model that was never watermarked.

By OpenAI's own description, a clean result can simply mean one of these:

• The passage is short. OpenAI's test caught about 80% of 200-token passages, not all of them.
• The wording was edited or translated after it was generated.
• The content is mathematics or similar, where OpenAI reports substantially lower detection.
• The text came from ChatGPT outside the EU, or from an API integration that never opted in.
• The text came from a different company's model.

The reverse limits matter just as much. OpenAI states that a watermark "does not measure human contribution," so it cannot tell a lightly prompted answer from a draft you reworked for an hour. It also "does not identify the user. It does not associate a person, organization, account, prompt, or conversation with the text."

Add OpenAI's own summary, "Text watermarking and detection remain early technologies with significant limitations," and the right reading is clear. A watermark result is one piece of evidence about origin. It is not a verdict on a person.

Why did OpenAI shelve its 2024 watermark and what changed?

OpenAI shelved its 2024 watermark because, by its own account, the method broke under large-scale rewording and could stigmatize AI use by non-native English speakers. What changed is the law. The EU AI Act's rule on machine-readable marking of AI text has applied since August 2, 2026.

In an August 2024 update, OpenAI called its earlier method "highly accurate and even effective against localized tampering, such as paraphrasing," but "less robust against globalized tampering; like using translation systems, rewording with another generative model," or inserting and deleting a special character between words. TechCrunch, citing the Wall Street Journal's 2024 reporting, adds that the company also feared users would move to rivals that did not watermark. The fairness worry is real on the style side too, and our page on why detectors misread non-native English explains it.

Then the rules arrived. Article 50(2) of the AI Act requires providers to ensure outputs "are marked in a machine-readable format and detectable as artificially generated or manipulated," and the European Commission lists these obligations as applicable from August 2, 2026. A later law, Regulation (EU) 2026/1744, gives systems placed on the market before that date until December 2, 2026. OpenAI's post does not say which date it is working to. It says only that the law requires this, which explains the EU-only launch.

Which AI models watermark text now?

Gemini, the newest Claude models and now OpenAI's models all watermark text in some form. Google DeepMind released SynthID-Text as open source on October 23, 2024 and runs it in Gemini. Anthropic began marking text from Claude models launched on or after August 2, 2026. OpenAI joined on October 5, 2026.

The three differ in reach. Google runs SynthID in Gemini's text output, covered in our Gemini SynthID explainer. Anthropic applies its mark worldwide to newer models, covered in our guide to Claude's text watermark. OpenAI is the narrowest so far: EU accounts first, API by opt-in.

Text is the late arrival at OpenAI. The company already adds C2PA Content Credentials to supported images and embeds SynthID watermarks in supported images and audio, with public verification for both.

What should you do about the ChatGPT watermark now?

The ChatGPT watermark changes little for most writers this week: it is not live for most accounts, nobody outside OpenAI can read it, and the checks you will meet still score style. The useful moves are to keep your drafts, know which account produced a text, and make the final wording your own.

1. Stop looking for hidden characters. The 2025 NNBSP scare was a formatting quirk. textGrain is in the words.
2. Do not trust a textGrain scanner. No public one exists.
3. Keep your version history. Notes, outlines and timestamped drafts show authorship to a person better than any watermark result.
4. Check how your text reads. Run it through our detector and look at the patterns it lists.
5. Rewrite, do not patch. If an AI-assisted draft you are allowed to use reads stiff, paste it into the humanizer and get it back in new wording from a model trained on real student essays. A free account covers 4 rewrites of 250 words each, one time, and paid plans start at $18 a month on the pricing page.

Where do these ChatGPT watermark facts come from?

Every third-party fact in this post was collected under the HumanizeMy Evidence Protocol v1.0 on October 6, 2026. The sweep identified 19 sources, set aside 4 (two could not be verified this run, one was off topic, one was a duplicate) and kept 15, which yielded 23 recorded items dated from January 24, 2023 to October 6, 2026. Two blind coding passes agreed on 96.2% of the labels for the first 21.

The kept sources are OpenAI's October 5, 2026 post and its textGrain technical report, OpenAI's August 2024 provenance update, three EU pages (the Article 50 text, the Commission's transparency page and Regulation (EU) 2026/1744), four arXiv papers (2301.10226, 2303.11156, 2303.13408, 2306.04634), three news reports, OpenAI's developer documentation and its October 5 post on X. The detection figures are OpenAI's own and have not been independently replicated. OpenAI says the rollout takes weeks, so this page will be re-read when the EU launch date is published.

Still unknown: the date watermarking starts for EU users, what makes an account "eligible," and the API setting and models involved, none of which OpenAI has documented.

Written by Fırat Mıhcı (ResearchGate). First published May 29, 2026, rewritten October 6, 2026 after OpenAI's announcement. No company named in this post pays us.