If you write, study, blog or run a business online, you have probably wondered whether AI text can be traced. In October 2026 OpenAI gave a concrete answer for the first time: a watermark that lives inside the wording itself. Readers cannot see it, and you cannot select or delete it like a logo on a photo.
This guide explains what OpenAI textGrain is, how it works in plain language, who actually gets it, how accurate it is according to reported figures, and, just as important, what it cannot prove. We also cover what it means for bloggers and site owners who use AI tools responsibly, including how Google treats AI-assisted content. Everything here is based on OpenAI’s own documentation and independent reporting, linked in the sources at the end.
- What Is OpenAI textGrain?
- Who Gets textGrain, and Since When?
- How textGrain Works, Step by Step
- How Accurate Is textGrain?
- What textGrain Cannot Prove
- Why the EU First? The AI Act Connection
- textGrain vs SynthID vs Claude’s Watermark
- What It Means for Bloggers, Students and Marketers
- Common Myths About AI Text Watermarks
- Frequently Asked Questions
- Sources
1. What Is OpenAI textGrain?
textGrain is OpenAI’s name for a text watermarking technology. A watermark on a photograph is a visible mark. A text watermark is different: the text looks completely normal, but the model’s word choices follow a hidden pattern created with a secret key. Anyone holding the matching detector can check a passage for that pattern.
OpenAI’s help centre describes the watermark as part of the wording itself, not something readers can see. The company says it built its own method, rather than using Google DeepMind’s SynthID or TextSeal, to get more control over the balance between how easy the watermark is to detect and how varied the model’s answers stay. OpenAI also says it plans to release textGrain as open-source technology so others can build on it.
In short, textGrain is a provenance signal. It helps answer the question “was this passage likely produced by an OpenAI model?” It does not answer “who wrote this?” or “is this true?”.
2. Who Gets textGrain, and Since When?
OpenAI announced textGrain on October 5, 2026. According to The Rundown’s coverage, it is being rolled out to eligible users of ChatGPT and Codex in the European Union over the following weeks, and API customers worldwide can turn it on for selected models, with the setting off by default.
| Product | Is textGrain on? | Where |
|---|---|---|
| ChatGPT text | Rolling out by default | European Union users |
| Codex text | Rolling out by default | European Union users |
| OpenAI API | Only if the developer opts in | Worldwide, selected models |
| ChatGPT outside the EU | Not a global default at this stage | Reported by independent analysts |
This matters because many headlines made it sound as though all ChatGPT text is now watermarked everywhere. Based on OpenAI’s own provenance page, text watermarks are listed as EU-only, while images and audio already carry other signals (Content Credentials and SynthID). Rollouts change quickly, so check OpenAI’s help page for the current status.
3. How textGrain Works, Step by Step
You do not need a maths background to understand the idea. A language model writes one small piece of text (a “token”) at a time, and at each step it has several good options with nearly equal probability. For example, after “The meeting was…” the words quick, fast and rapid may all fit.
Step 1: The model lists near-equal options
Whenever several words are almost equally likely, the model has a small amount of freedom. Watermarking uses exactly that freedom, so quality does not suffer.
Step 2: A secret key scores each option
textGrain combines a secret key with the preceding words to give every candidate a pseudo-random score. Anyone without the key sees what looks like ordinary randomness.
Step 3: The higher-scoring option is favoured
The model nudges its choice toward the candidate with the better score. One nudge means nothing. Across hundreds of words, the nudges add up to a pattern that would be extremely unlikely by chance.
Step 4: The detector recomputes the scores
To check a passage, the detector uses the same key, recalculates the scores for the words it finds, and sees whether the average is higher than chance would predict. If it is, the passage probably contains the watermark.
The analogy: imagine a writer who, whenever two words are equally good, secretly follows a private coin-flip rule only they know. Nobody reading the article notices. But someone who knows the rule can count how often the writer followed it and spot the pattern.
4. How Accurate Is textGrain?
Coverage of OpenAI’s tests reports a detector tuned for a target false positive rate of about 1%, meaning roughly one in a hundred human-written passages could be wrongly flagged. At that setting, one analysis of OpenAI’s reported results gives these detection rates:
| Scenario (as reported) | Approximate detection rate |
|---|---|
| Unedited passage of 200 tokens | About 80% |
| Unedited passage of 400 tokens | About 95% |
| 10% of words swapped | About 66% |
| 25% of words swapped | About 17% |
Two lessons follow. First, longer passages are far easier to detect: a few paragraphs give the detector much more evidence than a single sentence. Second, detection falls quickly as text is rewritten. OpenAI itself says substantial paraphrasing or translation can make a watermark undetectable. These figures come from OpenAI’s own evaluations as relayed by independent outlets, so treat them as lab results. OpenAI notes that strong lab performance does not guarantee reliable results in everyday use.
OpenAI also reports that the watermark caused no meaningful change in benchmark quality and does not identify the user. In other words, your prompts and identity are not embedded in the text.
5. What textGrain Cannot Prove
This section is the most useful one for anyone worried about false accusations. According to reporting based on OpenAI’s statements, a text watermark does not:
- Verify that the text is accurate. A watermarked answer can still be wrong.
- Show who owns the text. It says nothing about copyright or authorship rights.
- Measure how much a human contributed. A lightly edited draft and a pure copy-paste look the same to a simple detector, up to the point where edits erase the signal.
- Prove human authorship. The absence of a watermark does not mean a person wrote the text. It may come from another AI tool, or the watermark may have been weakened by editing.
- Cover other companies’ models. textGrain only flags text from OpenAI’s own systems.
WIRED reportedly relayed objections from developers worried about false positives and about light AI editing being treated as authorship. That is why teachers, editors and employers should never treat a watermark result as the only evidence in a dispute.
6. Why the EU First? The AI Act Connection
Analysts, including a Cloud Security Alliance research note, link the rollout to the transparency obligations of the EU AI Act, which asks providers of generative systems to mark synthetic content in a machine-readable way. OpenAI has said it wants to meet those rules, and independent reports say Anthropic made a similar move earlier, announcing text watermarking in August based on Google’s SynthID approach.
The practical result is a patchwork: text may be watermarked in one region and not another, and different companies use different methods. That is one reason experts say no single watermark can act as a universal “AI detector”.
7. textGrain vs SynthID vs Claude’s Watermark
| Feature | OpenAI textGrain | Google SynthID (text) | Anthropic (reported) |
|---|---|---|---|
| Type | Statistical, in word choice | Statistical, in word choice | Reportedly based on SynthID |
| Visible to readers | No | No | No |
| Who can detect it | OpenAI’s detector (not publicly checkable at launch) | Google’s tools | Vendor-specific |
| Survives heavy paraphrase or translation | No, weakens sharply | No, weakens sharply | Same general limit |
| Open-source plans | OpenAI says it plans to release it | Available in parts for developers | Not applicable |
The takeaway: all text watermarks share the same basic limits. They work best on long, unedited output from a single vendor’s model. If you want a broader view of how AI systems are built and used, our explainer on AI agents vs agentic AI is a good next read.
8. What It Means for Bloggers, Students and Marketers
For website owners and SEO
Google has repeatedly said it does not ban AI-assisted content. Its guidance, which was updated in October 2026, says AI can help with research and structure, while producing many pages with little added value can violate its scaled content abuse policy. Reporting on the update also says Google advises manually fact-checking AI-generated material before publishing. A watermark is not a ranking factor that anyone has confirmed. What matters is whether the page is accurate, original and useful.
A safe workflow looks like this:
- Use AI for research and outlines, not as the final author. Add your own experience, examples and opinion.
- Verify every fact, number and quote against a primary source before you publish.
- Edit for your own voice. Our free AI Text Humanizer can help smooth stiff phrasing, but it is an editing aid, not a substitute for fact-checking or for honesty about how you work.
- Be transparent where readers would reasonably want to know how a piece was made.
- Skip mass-produced pages. Ten careful articles beat a hundred thin ones.
For students and job seekers
Because a watermark can never prove authorship, schools and employers should not rely on it alone, and you should follow your institution’s AI policy. If you use AI to prepare a resume or cover letter, treat the output as a draft and rewrite it in your own words with real details. Our free ATS Resume Builder and the Career & Money AI Prompt Library are built for that kind of guided drafting.
For developers and teams
If you build on the API, textGrain is an opt-in feature, so you can decide whether provenance marking fits your product. Teams also need rules about who reviews AI output before it goes live. If you are planning automation, see our complete guide to AI automation agents, and use the DioxAI Prompt Lab to write clearer, safer instructions.
9. Common Myths About AI Text Watermarks
- Myth: “All ChatGPT text is now watermarked worldwide.” Reality: the rollout is aimed at the EU, with optional API use elsewhere.
- Myth: “Anyone can check a watermark with a free website.” Reality: at launch, the detector was described as not publicly checkable, so ordinary “AI detector” sites cannot read textGrain.
- Myth: “A watermark proves a student cheated.” Reality: OpenAI itself says it cannot prove authorship or the share of human contribution.
- Myth: “Watermarked text ranks worse on Google.” Reality: there is no confirmed link. Google’s stated focus is helpfulness and quality.
- Myth: “No watermark means a human wrote it.” Reality: other tools, other regions, or editing can all mean no signal is found.
10. Frequently Asked Questions
What is OpenAI textGrain?
It is an invisible statistical watermark that OpenAI announced on October 5, 2026. It hides a pattern in the model’s word choices so a detector can estimate whether text came from OpenAI’s models.
Does ChatGPT watermark all text now?
Not everywhere. The watermark is rolling out for ChatGPT and Codex text in the European Union, and API customers worldwide can opt in on selected models. It is not described as a global default.
Can I see or copy the watermark?
No. It is part of the word choices, not a visible mark or hidden character, so there is nothing to select or delete.
How accurate is textGrain?
Reported results show about 80% detection on 200-token passages and about 95% on 400 tokens at a 1% false positive rate. Accuracy drops quickly when many words are changed, and OpenAI says paraphrasing or translation can make the watermark undetectable.
Can textGrain prove that a student or writer used AI?
No. It gives a probability that text came from OpenAI’s system. It cannot confirm accuracy, ownership, how much a human contributed or who wrote the text, so it should never be the sole evidence.
Will textGrain hurt my Google rankings or AdSense approval?
There is no confirmed ranking effect from watermarks. Google’s guidance focuses on whether content is helpful, original and accurate, and warns against mass-produced pages. Fact-check and add real value to anything you publish.
Is textGrain the same as SynthID?
No. Both are statistical watermarks, but OpenAI says it developed its own method for more control over detectability and response variety. OpenAI also uses SynthID signals for images and audio.
Will textGrain be open source?
OpenAI says it plans to release textGrain as open-source technology. Check OpenAI’s announcements for the current status.
Conclusion
OpenAI textGrain is a meaningful step toward labelling AI-generated text, but it is a narrow tool. It is invisible, probabilistic, mostly EU-focused for now, and easy to weaken with heavy rewriting. It cannot settle questions of truth, ownership or authorship. For creators, the lesson is simple and old-fashioned: use AI as a helper, check your facts, add your own experience, and publish things you would be proud to put your name on. That approach protects your readers, your Google visibility and your long-term credibility.
Want to keep learning? Read our guides on AI agents vs agentic AI and AI automation agents, or plan your next article with the free Mind Map Maker.
Sources
- OpenAI Help Center: Provenance signals in OpenAI-generated content
- The Rundown: OpenAI plans textGrain watermarks for ChatGPT and Codex in the EU
- Cloud Security Alliance: OpenAI textGrain, watermarking text under the EU AI Act
- Slashdot, citing The Verge: OpenAI is adding text watermarking in ChatGPT and Codex
- Pasquale Pillitteri: OpenAI unveils textGrain (source of the reported detection figures)
- Google Search Central: guidance on generative AI content (updated October 2026). Read the original on Google’s documentation site to confirm the current wording.
Mansoor is a web developer, AI builder, and SEO specialist who turns ideas into tools people actually use. Through DioxAI, he builds free browser-based tools and writes practical guides for job seekers and creators, blending real development work with a sharp eye for search visibility.