# ChatGPT Text Watermark Explained: What textGrain Means for Your Writing - Link: https://thejusgrow.com/chatgpt-text-watermark-explained/ - Published: 2026-10-14T10:37:45+00:00 - Author: Varun Sharma ChatGPT text watermark: the short answer The ChatGPT text watermark, called textGrain, is an invisible statistical pattern in the words ChatGPT chooses. OpenAI is adding it to eligible ChatGPT and Codex text in the EU over the coming weeks. Meanwhile, API customers anywhere can opt in. However, only approved researchers can run the detector, and a match does not prove who wrote something. On 5 October 2026, OpenAI announced that a ChatGPT text watermark would start appearing in the European Union. The change follows from the EU AI Act’s rule that AI-generated text should be identifiable by machines. So if you live in the EU, the text you copy out of ChatGPT may soon carry a signal you cannot see. That naturally raises questions. For example, can your teacher check it, and does it reveal who you are? This explainer answers those questions using [OpenAI’s announcement](https://openai.com/index/eu-text-provenance/), its help centre and its technical report. TheJusGrow has not tested the ChatGPT text watermark, so every figure below is OpenAI’s own unless we say otherwise. ## What is the ChatGPT text watermark? In short, textGrain is OpenAI’s method for nudging the model’s word choices in a way only a secret key can reveal. When ChatGPT picks between several reasonable next words, the watermark tilts that choice slightly. Then, across a long passage, those tiny tilts add up to a pattern a detector can spot. Importantly, your copied text carries nothing extra. According to [OpenAI’s help centre](https://help.openai.com/en/articles/8912793-provenance-signals-in-openai-generated-content), textGrain adds no hidden characters or extra tokens and doesn’t change copied text. In other words, a ChatGPT text watermark lives in which words appear, not in invisible symbols. OpenAI also says it plans to open-source textGrain, although that has not happened yet. This is also different from the image labels many people know. For example, our guide to the [SynthID Detector](https://thejusgrow.com/synthid-detector/) covers Google’s tool for checking watermarked images, video and audio. ## Why is OpenAI adding it now? The trigger is European law. In particular, Article 50 of the EU AI Act requires providers to mark AI-generated content in a machine-readable way, and those duties apply from 2 August 2026. Meanwhile, the European Commission published a final Code of Practice on marking and labelling AI content on 10 June 2026, as [law firm Jones Day reported](https://www.jonesday.com/en/insights/2026/06/european-commission-publishes-final-code-of-practice-on-marking-and-labelling-aigenerated-content). The Code itself is voluntary, but it still gives companies a recognised way to meet the law. That is why OpenAI is rolling the watermark out in the EU first, rather than switching it on everywhere. ## How does textGrain work, in plain English? First, it helps to know that language models write by predicting one word at a time from a list of likely options. Usually several options are almost equally good. So textGrain uses a secret key to decide which of those near-equal options to favour, so the final text carries a faint, consistent bias. OpenAI’s [technical report](https://cdn.openai.com/pdf/e9508624-d767-41b6-a26d-e34ca798ada6/textgrain-entropy-calibrated-watermarking-for-language-model-text.pdf), written with academic co-authors, also adds a safeguard. It caps how much randomness the watermark gives up, so writing quality should stay close to normal. As a result, averaged across keys, the output matches what the model would produce anyway. The detector, therefore, only needs the text and the key, not your account or conversation. OpenAI reports almost no quality cost. For instance, on one reasoning benchmark, GPT-6 Astra scored 94.44% without the watermark and 93.94% with it. OpenAI also says thumbs-down rates and speed showed no meaningful change, although no independent group has confirmed these results yet. ## Who does the ChatGPT text watermark affect? The scope is narrower than many headlines suggest. As of October 2026, the ChatGPT text watermark applies as follows: | Situation | Watermarked? | Notes | | ChatGPT in the EU | Rolling out over the coming weeks | Covers all EU plans, but not yet every user | | Codex in the EU | Rolling out as well | Also covers eligible text output | | ChatGPT outside the EU | No | Not a global default, since the rule is European | | OpenAI API, anywhere | Only if you opt in | Otherwise off by default | | Short replies under about 150 words | Not required | Exempt when under 200 tokens | | Code snippets | Not required either | Also exempt under the Code | Developers who want it can then switch it on in their API project or organisation settings, under Text provenance. According to OpenAI, older models will follow over the coming weeks. However, none of the sources we checked mention an opt-out for EU ChatGPT users, so we can’t say whether one exists. ## Who can check for the watermark? Very few people, at least for now. In fact, OpenAI says the text detector is available only to approved academic and research organisations, which apply case by case. Access is free, but there is no public text checker. OpenAI’s public verification page, meanwhile, handles only images and audio. OpenAI also rules out several uses. For example, the detector isn’t offered for commercial fraud prevention, authentication, resale or checking student work. So a teacher can’t upload an essay to OpenAI and ask whether ChatGPT wrote it. Our guide to [AI agents for teachers](https://thejusgrow.com/ai-agents-for-teachers/) covers what schools can use instead. ## How reliable is the ChatGPT text watermark? Reliability depends heavily on length, editing, subject and language, so the ChatGPT text watermark is far from foolproof. OpenAI published these results at a 1% false- positive target. That means about one in a hundred human texts could still be flagged wrongly: | Test | Detection rate | | Unedited 200-token passage | About 80% | | Unedited 400-token passage | About 95%, so length helps | | 400 tokens, 10% of words swapped for synonyms | Down from about 92% to 66% | | 400 tokens with 25% of words swapped | About 17% | | Spanish text | 69.0%, although English figures are higher | | Romanian text | Only 42.2% | Maths content performs much worse, because there are fewer interchangeable words to work with. OpenAI also says it strengthened the watermark for languages that scored below 60%. Even so, [The Register](https://www.theregister.com/ai-and-ml/2026/10/06/openai-rolls-out-weak-sauce-watermarking-for-ai-text/5301257) described the protection as weak, given how quickly light editing erodes it. The takeaway is simple: a ChatGPT text watermark is a signal about some long, lightly edited passages, not a reliable fingerprint for every piece of AI writing. ## What a detection does and doesn’t prove OpenAI is unusually direct about the limits. Specifically, according to its help centre, a positive result does not: * Measure how much a human edited the text. * Establish ownership or responsibility, either. * Identify the user, account, prompt or conversation. * Check whether the content is accurate, since accuracy is a separate question. The reverse is also true, because a missing watermark doesn’t prove a human wrote something. Short, edited or translated text may slip through, and so may older models and other companies’ tools. Therefore, nobody should treat any single detector result as proof of cheating or innocence. ## ChatGPT vs Claude: how the two watermarks compare OpenAI is not alone, though. In August 2026, [Anthropic explained](https://www.anthropic.com/news/claude-text-watermark) that Claude applies a version of Google DeepMind’s SynthID-Text worldwide, because it lacks a durable way to limit it by region. | | ChatGPT (textGrain) | Claude | | Method | OpenAI’s own textGrain | A version of SynthID-Text instead | | Where it applies | EU only, for now | Worldwide | | API use | Opt-in, off by default | Not stated | | Who can detect | Approved researchers | Eligible organisations in a private preview | | Public checker | None for text | None announced so far | | Carries user identity | No, according to OpenAI | Also no, according to Anthropic | Because each company uses its own method and key, no single detector covers all AI text. That is worth remembering whenever a website claims it can spot any AI writing. ## What the ChatGPT text watermark means for you ### Students and parents First, follow your school’s rules on AI help, and keep drafts and notes as evidence of your own process. Since OpenAI’s detector isn’t offered for checking student work, a watermark alone shouldn’t decide an academic case. Our guide to [AI agents for students](https://thejusgrow.com/ai-agents-for-students/) also covers fair ways to use AI for study. ### Workers and freelancers If you are in the EU, long and lightly edited ChatGPT output may carry a mark. So disclose AI help wherever a client or employer requires it, rather than assuming nobody can tell. ### Businesses and developers According to OpenAI, the ChatGPT text watermark doesn’t replace a visible AI label. So businesses publishing AI text in the EU should still check their Article 50 labelling duties. Developers, meanwhile, need to decide whether to turn on Text provenance in the API. For wider safety habits, also see our guide on whether [AI agents are safe](https://thejusgrow.com/are-ai-agents-safe/). ## Checklist: living with the ChatGPT text watermark * Know your location, because the watermark currently applies to EU ChatGPT and Codex output. * Assume long, lightly edited EU output may carry a mark, even after small edits. * Disclose AI help whenever your school, client or employer requires it. * Never treat a detector result, positive or negative, as proof on its own, because both can mislead. * Keep drafts and notes, since they show your own process. * Businesses should also review visible labelling duties under Article 50. * Finally, developers can decide whether to enable Text provenance for their API projects. ## What to watch next Several open questions will shape how much the ChatGPT text watermark matters. First, OpenAI says it plans to open-source textGrain, which would let outside researchers test the method themselves. Second, the EU rollout is gradual, so it is not yet clear when every EU account will be covered. Third, OpenAI says older API models will be added over the coming weeks. Finally, all accuracy and quality figures so far come from OpenAI. Independent tests, once they appear, will show whether the reported detection rates hold up in real-world writing. Until then, treat the numbers in this article as company claims rather than settled facts. Key takeaways * textGrain is an invisible pattern in ChatGPT’s word choices, not hidden characters. * It is rolling out to eligible ChatGPT and Codex text in the EU, while API users anywhere can opt in. * Only approved researchers can use OpenAI’s text detector, so teachers cannot check essays with it. * However, detection drops sharply with short text, heavy editing, maths and some languages. * A match proves neither authorship nor accuracy, and a missing mark proves nothing either. ## ChatGPT text watermark: FAQs **Does ChatGPT watermark my text?** Only in the EU for now, and OpenAI says the rollout will take place over the coming weeks. Outside the EU, API users can opt in, but ChatGPT itself isn’t watermarked by default. **Can my teacher or employer check for it?** Not through OpenAI. Only approved researchers can use the text detector, and OpenAI also says checking student work is out of scope. **Does editing weaken the watermark?** Yes. For example, in OpenAI’s tests, swapping 25% of words for synonyms cut detection on 400-token passages to about 17%. **Does the watermark reveal who I am?** No. According to OpenAI, a ChatGPT text watermark detection doesn’t identify the user, account, prompt or conversation. ## Sources * [OpenAI, “Our approach to EU text provenance rules” (5 October 2026)](https://openai.com/index/eu-text-provenance/) * [OpenAI Help Center, “Provenance signals in OpenAI-generated content” (checked 13 October 2026)](https://help.openai.com/en/articles/8912793-provenance-signals-in-openai-generated-content) * [The Register, report on OpenAI’s text watermarking (6 October 2026)](https://www.theregister.com/ai-and-ml/2026/10/06/openai-rolls-out-weak-sauce-watermarking-for-ai-text/5301257) * [OpenAI and co-authors, “textGrain: Entropy-Calibrated Watermarking for Language Model Text” (5 October 2026)](https://cdn.openai.com/pdf/e9508624-d767-41b6-a26d-e34ca798ada6/textgrain-entropy-calibrated-watermarking-for-language-model-text.pdf) * [Anthropic, “How Claude’s text watermarking works” (14 August 2026, updated 1 September 2026)](https://www.anthropic.com/news/claude-text-watermark) * [Jones Day, “European Commission publishes final Code of Practice on marking and labelling AI-generated content” (June 2026)](https://www.jonesday.com/en/insights/2026/06/european-commission-publishes-final-code-of-practice-on-marking-and-labelling-aigenerated-content)