OpenAI has started marking its own words. On October 5, 2026, the company announced it will add an invisible watermark to text generated by ChatGPT and Codex in the European Union, a move driven by the EU AI Act’s transparency rules. The system is called textGrain, and it works by quietly shaping the model’s word choices so that software built for the job can recognize AI-generated text, even though readers will never see a thing.
Direct answer: OpenAI textGrain is an invisible statistical watermark that OpenAI is rolling out to ChatGPT and Codex text in the EU to comply with the EU AI Act. It steers the model’s word choices into a detectable pattern instead of adding hidden characters or metadata, so the signal survives copy and paste. EU users will get it automatically in the coming weeks, while API developers worldwide can switch it on as an option starting now. Detection depends on text length and light editing can sharply weaken it.
Why OpenAI is watermarking ChatGPT text now
The timing is regulatory, not optional. The EU AI Act’s transparency obligations, which took effect on August 2, require providers of AI systems to make their generated content machine-identifiable. OpenAI said Monday that textGrain is its answer: eligible ChatGPT and Codex text output produced in the European Union will begin carrying the watermark automatically over the coming weeks, across all plans, with no off switch.
Outside Europe the picture is different. OpenAI is not making text watermarking a global default at launch. Developers using OpenAI’s API anywhere in the world can turn it on for select models starting today, but it is off by default. The company says cloud partners like Microsoft Azure will offer it in the coming weeks.
If you use ChatGPT in the United States, the United Kingdom, or Canada, this change does not touch your chats by default. But it still matters to you: watermarking changes the provenance of text your employees, freelancers, students, and vendors may be handing you, and it signals where detection technology is heading.
How textGrain actually works
Forget hidden characters, zero-width spaces, or secret metadata. TextGrain lives inside the words themselves. OpenAI describes the method this way: as the model generates text, it uses a secret key to rank the most likely next words and subtly favors certain choices over others. One nudge means nothing. Hundreds of nudges across a passage build up a statistical pattern that special software holding the same key can recognize.
OpenAI developed textGrain with researchers from the University of Pennsylvania and Yale, and published a technical report alongside the announcement walking through the method with a worked example.
Two properties matter most for readers:
- It is invisible and durable. There is nothing to see, nothing to delete, and no metadata to strip. Copying the text into a document, an email, or a website keeps the pattern intact.
- It identifies the text, not the person. OpenAI says the watermark does not identify the user who generated it, cannot reveal prompts, and does not expose chat history. The company also reports no meaningful change in model performance with textGrain switched on.
Who can see the watermark: detectors and access
The watermark is machine-readable, but the machines that can read it are tightly controlled at launch. OpenAI will not make its detector public. Instead, access goes to approved researchers and expert organizations on a case-by-case basis, so they can evaluate how reliable the system is and where responsible use makes sense. The company says it plans to release the textGrain technology as open source later.
That means teachers, hiring managers, and editors should not treat textGrain as an off-the-shelf plagiarism checker today. A positive detection can indicate that OpenAI tooling likely generated or processed part of a passage, but OpenAI is explicit that a negative result proves nothing: the text might be too short, heavily edited, translated, or produced by another company’s model.
For a broader sense of how the AI landscape is shifting for businesses, see our comparison of the two leading chatbots https://techtorev.com/claude-vs-chatgpt-small-business-2026/ and our rundown of OpenAI’s always-on agents https://techtorev.com/openai-dots-explained-price-release-date-how-it-works/
How accurate is detection, and what breaks it
OpenAI published its own stress tests, and the numbers are honest about the limits. In one test, replacing just 10% of the words in a passage with synonyms dropped detection from about 92% to 66%. Independent analysis put detection at roughly 80% for a 200-token passage and about 95% at 400 tokens, with a 1% false-positive rate. Replacing a quarter of the words with synonyms cut detection to around 17%.
Short passages are harder to catch, because the statistical signal needs length to accumulate. Mathematical answers and translated text are also harder to detect, since constrained wording gives the model fewer free choices to nudge. The practical upshot: textGrain is a compliance tool for provenance, not a guarantee that AI writing will always be caught. Anyone who lightly rewrites AI output can weaken the signal substantially.
TextGrain vs Claude and Google watermarking
OpenAI is not the first mover here. Anthropic launched a machine-readable marker for Claude back in August 2026, and its approach is stricter: Claude’s output is watermarked everywhere, with no opt-out. Google DeepMind has been watermarking its model outputs for a couple of years through its SynthID system. (Oddly, Google moved the other direction in August, announcing that users could remove watermarks from generated images, videos, and songs in countries where they are not required.)
The contrast is telling. Anthropic applies watermarking globally and mandatory; OpenAI applies it mandatorily only where the law demands and makes it optional everywhere else. That divergence gives users and businesses a real choice between vendors with different provenance policies, and it means the same AI-written text may be detectable or not depending on which model produced it.
What this means for students, freelancers and small businesses
For most American small businesses, the immediate effect is indirect but real. If you hire freelancers or accept written deliverables from EU-based contractors, the text you receive may soon carry an invisible marker that future detectors can flag. Contract terms about original work and disclosure of AI assistance are worth revisiting now, not after a dispute.
For students and teachers, the message is narrower than it sounds: there is no public detector, and light editing degrades the signal, so textGrain does not end the AI-cheating debate. It does, however, give institutions with researcher access a real tool for auditing large bodies of text.
For businesses choosing AI vendors, watermark policy is now a purchasing consideration. If provable human authorship matters in your industry, ask your vendor what provenance controls it offers and whether they are on or off. And if AI pricing is part of your stack decision, our breakdown of Google’s latest pricing changes https://techtorev.com/free-gemini-flash-lite-october-9-explained/ shows where the market is moving, while our look at Meta’s new assistant https://techtorev.com/meta-muse-for-small-business-explained/ covers the competing option many small businesses are testing.
Frequently asked questions
What is OpenAI textGrain?
TextGrain is OpenAI’s invisible statistical watermark for AI-generated text. It subtly shapes the word choices ChatGPT and Codex make, creating a machine-readable pattern that detectors can spot. OpenAI announced it on October 5, 2026, and built it with researchers from the University of Pennsylvania and Yale.
Is ChatGPT adding a watermark to all text?
No. The watermark becomes mandatory only for eligible ChatGPT and Codex users in the European Union, rolling out over the coming weeks to comply with the EU AI Act. API users worldwide can opt in starting now, but it is off by default and OpenAI is not making it a global default at launch.
Can readers see the ChatGPT watermark?
No. It is invisible. There are no hidden characters, symbols, or metadata. The watermark is embedded in the statistical pattern of the words themselves, so it survives copying and pasting.
Can the textGrain watermark be removed by editing?
It can be weakened. OpenAI’s tests show replacing 10% of words with synonyms drops detection from about 92% to 66%. Short text, math content, and translated text are also harder to detect. OpenAI says the detector cannot identify who wrote the text or reveal prompts.
How is textGrain different from Claude’s watermarking?
Anthropic’s watermarking, launched in August 2026, applies to Claude’s output everywhere with no opt-out. OpenAI’s is mandatory only in the EU under the AI Act and optional elsewhere. Google DeepMind’s SynthID has been watermarking outputs for a couple of years.
Is there a public ChatGPT watermark detector?
Not yet. OpenAI is limiting detector access to approved researchers and expert organizations at launch, citing the risk of false positives, and plans to open source the technology later.