What Is the New ChatGPT Text Watermark?: What textGrain Marks, and What a Public Checker Still Cannot See

What OpenAI actually turned on

OpenAI's own note, Our approach to EU text provenance rules, is the document to trust over a screenshot of a setting that does not exist yet. Starting October 5, API customers anywhere can opt in to watermarked text for select models, and the switch stays off until they choose it. Over the coming weeks, eligible ChatGPT and Codex users in the European Union, on every plan, are supposed to get the watermark on their text output. OpenAI said it is not making text watermarking a global default at launch.

The reason for the EU timing is the AI Act's rule that generated text should be identifiable in a machine-readable way. Ars Technica's report puts the same limit in plainer language: the watermark is the default only in the EU, it is optional elsewhere, and it is not a reliable stamp. A chat opened outside that rollout, an older transcript, or an API call that never opted in can be ChatGPT text with no watermark in it at all.

Images and audio were already on a different system. OpenAI has pointed people to Content Credentials and SynthID-style marks for those formats, plus a public verification page. Text did not get that public check. If someone tells you to upload an essay to openai.com and read a watermark badge, they are describing the image and audio tools, not textGrain.

How an invisible watermark sits in ordinary words

A language model picks the next piece of text from a list of likely options. textGrain, as OpenAI describes it, adds an invisible statistical signal to those choices, tied to a secret so the company's detector can ask whether a passage still carries an OpenAI watermark. It does not insert zero-width spaces, odd punctuation, or a code you can highlight. The sentences are still sentences.

Because the signal lives in which words were preferred, a clean copy and paste keeps it. Translation, a heavy rewrite, a short answer, or a math solution with almost no flexible wording does not. OpenAI's technical report on textGrain is the long version of that design. The public summary is enough for a writing decision: the mark is a pattern in likely words, and patterns in likely words are exactly what a paraphrase disturbs.

OpenAI also says the watermark does not name the person, the account, the school, the prompt, or the conversation. A positive hit, if you could even get one, would be "this passage still looks watermarked," not "this student pressed the button." A miss would mean even less.

Who is allowed to look for the mark

OpenAI opened applications on October 5 for a detector that reports whether it sees an OpenAI watermark. Access starts with approved researchers and expert organizations, granted case by case, because the company does not want a public tool that misses marks and also invents them. That is the opposite of a website where anyone pastes a paragraph and gets a yes or no.

The numbers they did publish are narrow. At a target false-positive rate of 1%, the detector found watermarks in about 80% of 200-token passages and about 95% of 400-token passages for flexible prose such as psychology, and the rate was substantially lower for mathematics, where the model has fewer free word choices. In a test on 400-token passages, replacing 10% of the words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%. Those are OpenAI's figures, under their conditions, not a promise about your assignment.

A campus report is not this detector unless that campus has been given the key, and OpenAI has not said that it has. What AI detector Turnitin uses is still a separate classifier that scores how generated the writing looks. Does Turnitin detect ChatGPT is the page for that badge. textGrain does not replace it, and a kind result on a public checker does not cancel it.

A public writing score is not a watermark check

DetectGPT does not hold OpenAI's secret, and it does not claim to. It reads the finished draft the way other public checkers do: even sentences, safe wording, flat rhythm, then a percentage and the lines that still look generated. That can fire on ChatGPT, Claude, Grok, Gemini, or on a person who writes in a very tidy voice. It can also stay quiet on watermarked text that was rewritten until the rhythm looks human, and on unwatermarked text that still sounds like a model.

Treat the two results as answers to two questions. textGrain, when it exists on a passage and when someone with access runs the restricted detector, asks whether OpenAI's mark is still in the words. A style score asks whether the page sounds machine-made. How to read an AI writing score is the second question, including why 40% can be two stiff paragraphs rather than a confession.

If you want that second check, paste the draft on DetectGPT or use the URL tab on a public page. The floor is 50 words, the cap is 3,000, and there is no account. Do not describe the percentage as "watermarked" or "not watermarked." That word now belongs to a specific OpenAI signal most people cannot query.

What an edit does to the signal, without a costume

OpenAI's synonym test is the useful caution, not a recipe. Swapping a tenth of the words already knocked their detection down by a wide margin, and swapping a quarter of them nearly emptied it. A real revision, the kind where you rebuild a sentence because you know the source, will move the wording even more. So will a translation. So will cutting a long answer down to a caption. The watermark was never sold as something that survives ordinary editing.

The same edit can leave the style problem sitting there. Synonym swaps keep the skeleton of the sentence, which is why a spinner often still looks generated to a rhythm checker even after a watermark would have failed. If the goal is a paragraph you can defend, write the point the way you would say it, keep the facts, and keep a dated draft. DetectGPT's AI humanizer can loosen a stiff passage up to 1,000 words, and it does not promise that textGrain, a campus indicator, or any other detector will agree with the result.

Absence is not proof either. OpenAI says a missed watermark can mean the text was short, edited, translated, from a model that was not covered, from before this rollout, or from another company's tool. Human writing has no textGrain mark because it was never generated under that key. Those two situations look the same to a detector that only hunts for the mark.

FAQs

What is the new ChatGPT text watermark?

It is textGrain, an invisible statistical pattern OpenAI adds by nudging the model's word choices. Readers do not see a symbol, and a plain copy of the reply keeps the pattern because the pattern is the wording.

Is every ChatGPT answer watermarked now?

No. As of the October 5, 2026 announcement, ChatGPT and Codex watermarking is planned for eligible users in the EU over the coming weeks. It is not a global default. API watermarking is opt-in and off unless the customer enables it.

Can I check the ChatGPT watermark myself?

Not with a public tool. OpenAI is giving detector access to approved researchers and expert organizations, case by case. A paste box on the open web is not that detector.

Does DetectGPT detect textGrain?

No. DetectGPT scores how machine the writing sounds. It does not have OpenAI's watermark key, and a percentage here is not a watermark present or watermark absent result.

Does the watermark identify the student or the account?

OpenAI says no. The mark does not attach a person, organization, account, prompt, or conversation to the text.

If no watermark is found, is the writing human?

No. The text may never have been watermarked, or it may have been shortened, edited, or translated until the signal failed. OpenAI says a miss does not prove human authorship.

Will a small paraphrase remove it?

OpenAI's evaluation on 400-token passages dropped from about 92% detection to 66% after 10% of the words were replaced with synonyms, and to 17% after 25% were replaced. That is a limitation of the mark, not a guarantee about any other detector.