Is My ChatGPT Text Watermarked Now? What OpenAI’s textGrain Means for EU Users, Students, and API Developers (2026)

A hidden signal in the words themselves. That’s what OpenAI started shipping this week, and it has a lot of people wondering whether the essay or client report they wrote with ChatGPT is now “tagged.”

For most readers outside Europe, the answer is no. For EU users it’s yes, or it will be within weeks. Either way, the watermark is far less of a lie detector than the headlines make it sound.

Here’s who is affected, how well detection actually works, and what you should do about it (mostly: disclose honestly and stop worrying about the wrong thing).

Jump Through This Explainer

• Is My ChatGPT Text Watermarked Now?
• How Does OpenAI’s textGrain Watermark Work?
• How Reliable Is textGrain Detection?
• Four Things the Watermark Can’t Tell Anyone
• Should Students and Freelancers Worry About the ChatGPT Watermark?
• Should API Developers Turn textGrain On?
• Myths Already Spreading
• Tips
• Warnings
• FAQ
• Sources

Is My ChatGPT Text Watermarked Now?

ChatGPT and Codex text is being watermarked only for users in the European Union, where OpenAI is rolling out textGrain to all plans over the weeks after its October 5, 2026 announcement, according to The Next Web’s report on the launch.

Outside the EU, ChatGPT and Codex output is not watermarked under this plan. The API is a separate case: since October 5, developers worldwide can switch textGrain on for selected models, but it stays off by default.

Table: our summary of OpenAI’s rollout as reported by The Next Web and The Decoder, checked October 8, 2026.

That last row matters more than people think. If a writing tool you pay for runs on OpenAI’s API, whether its output is marked is the tool maker’s decision, not yours.

How Does OpenAI’s textGrain Watermark Work?

OpenAI’s textGrain watermark works by adding a hidden statistical signal to the model’s word choices, so a matching detector can later spot the pattern without any visible mark, extra characters, or metadata in the text.

Think of it as a slight, consistent preference among words that would all fit. A human reader sees normal prose. A detector that knows the pattern sees a skew that’s very unlikely to happen by chance.

A few practical consequences follow from that design:

• You can’t see it. There’s nothing to highlight or delete. It isn’t hidden Unicode characters, which is a different thing some people have reported in copied AI text.
• It needs length. A short reply doesn’t contain enough word choices to carry a reliable signal.
• It lives in the wording. Change enough words and the signal fades. Translate it and the original choices are gone.

OpenAI also says the watermark made no meaningful difference to benchmark scores, and it plans to release the technique as open source.

So no, ChatGPT in Paris isn’t going to write worse than ChatGPT in Chicago.

How Reliable Is textGrain Detection?

OpenAI reports that textGrain detects about 95% of 400-token ChatGPT passages and about 80% of 200-token passages at a 1% false-positive rate, according to figures published with the October 2026 launch and summarized by The Decoder.

Those are the headline numbers. The fine print is where it gets interesting. When I compared the full set of figures, I found that detection falls off a cliff once text is edited, and it’s weaker on some subjects than others.

Table: detection rates OpenAI published at launch, as reported by The Decoder and The Next Web (checked October 8, 2026). Token-to-word conversion is approximate.

Read the bottom two rows again. The watermark is good at catching copy-paste. It’s poor at catching someone who rewrote the draft.

That’s by design, not a flaw. A watermark answers “did this exact text come out of the model?” It was never going to answer “did a person use AI somewhere in the process?”

Four Things the Watermark Can’t Tell Anyone

The textGrain watermark cannot identify which user generated a piece of text, measure how much a human contributed, establish who owns the text, or confirm whether it’s accurate, according to OpenAI’s own description of the system.

• Who wrote the prompt. The detector doesn’t reveal your account or your prompts.
• How much is yours. It can’t say “60% human.” It flags a signal or it doesn’t.
• Ownership. A watermark isn’t a copyright claim by OpenAI.
• Truth. Watermarked text can be correct. Unwatermarked text can be wrong.

And the one that matters most for anyone accused of using AI: a missing watermark does not prove a human wrote the text. It could be edited, translated, written by another model, or written outside the EU where marking is off.

I noticed that several early reaction posts skip this point entirely. It cuts both ways. The watermark can’t clear you, and on edited text it can’t convict you either.

Should Students and Freelancers Worry About the ChatGPT Watermark?

Students and freelancers in the EU should assume that unedited ChatGPT text longer than about 300 words is detectable by textGrain, but as of October 2026 their teachers and clients can’t run that check, because OpenAI limits the detector to approved researchers and expert organizations.

Access is granted case by case under the EU’s transparency Code of Practice, per The Next Web. There’s no public checker, no browser plugin, and no plagiarism-tool integration announced.

Honestly, that’s the less important part. Policies on AI use already exist whether or not anyone can detect it.

• Students: your school’s AI policy is the rule that counts. If it requires disclosure, disclose. A watermark doesn’t change what was allowed yesterday.
• Freelancers: check your contracts. Some clients now ask whether deliverables used AI. Answering truthfully is cheaper than being caught by a tool that arrives next year.
• Content writers: if a client bans AI drafting, the watermark is the least of your problems. Talk to them about what’s allowed.

If you’ve built income around AI-assisted writing, our piece on what OpenAI’s custom GPT retirement means for AI side hustles covers another platform change worth planning for.

Should API Developers Turn textGrain On?

API developers outside the EU are not required to enable textGrain, which OpenAI made available as an opt-in for selected models on October 5, 2026, but providers whose generative AI output reaches EU users should review their marking duties under Article 50 of the EU AI Act.

Article 50 of the AI Act requires providers of systems that generate synthetic text, audio, images, or video to mark the output in a machine-readable, detectable format. It also carves out tools that only perform an assistive editing function and don’t substantially change the input.

A rough way to decide:

1. Map where your users are. If a meaningful share are in the EU and your app generates new text for them, talk to counsel about Article 50.
2. Check what your app actually does. A grammar fixer is in a different position from an essay generator.
3. Look at the European Commission’s guidance. The Code of Practice on transparency of AI-generated content is the document regulators will point to.
4. Test output quality with it on. OpenAI says benchmarks didn’t move. Your own prompts are still worth a quick before-and-after.

This isn’t legal advice. It’s a checklist for the conversation you probably need to have.

Myths Already Spreading

A lot of confident advice appeared within hours of the announcement. Some of it is wrong.

• “Use a VPN to avoid it.” Rollout is tied to how OpenAI determines you’re an EU user, and OpenAI hasn’t published that logic. Don’t build a plan on guesses. More to the point, dodging the mark doesn’t change your school’s or client’s rules.
• “Turnitin can detect it now.” No public integration has been announced. The detector is restricted.
• “It identifies you.” OpenAI says the detector checks for the signal without identifying users or revealing prompts.
• “No watermark means human-written.” Covered above. It doesn’t.

When I checked the wording in OpenAI’s launch coverage against these claims, the gap was obvious: almost every viral claim assumes a public detector that doesn’t exist yet.

If you’re also tightening up account safety while you’re in ChatGPT’s settings, see how to check an unknown sign-in in your ChatGPT security history.

Tips

• If you’re in the EU, treat any long, unedited ChatGPT passage as marked from now on.
• Keep your drafts and version history. If authorship is ever questioned, a document history beats any detector.
• Read your institution’s or client’s AI policy this week, not after a dispute.
• Developers: log whether textGrain is on per project, so you can answer a compliance question later.

Warnings

• Don’t pay for “watermark remover” tools. They can’t verify they did anything, and they don’t change what your policy allows.
• Don’t treat a detector result, from any vendor, as proof. OpenAI itself says a missing mark proves nothing, and editing weakens a present one.
• Rollout dates are “coming weeks.” Your account may not have it yet, or may already.
• Third-party AI checkers that claim to read OpenAI’s watermark are guessing. The real detector isn’t public.

FAQ

Does the ChatGPT watermark apply to text I write myself and paste into ChatGPT?

No. textGrain shapes the words the model generates, so your own text isn’t marked. If ChatGPT rewrites your text, the rewritten output can carry the signal.

Is ChatGPT text watermarked in the United States?

Not under the plan OpenAI announced on October 5, 2026. Automatic watermarking covers ChatGPT and Codex users in the EU, while API customers anywhere can choose to switch it on.

Can I check if my own text has an OpenAI watermark?

Not yet. OpenAI limits its textGrain detector to approved researchers and expert organizations, reviewed case by case. There is no public checker.

Sources

• The Next Web — OpenAI starts watermarking ChatGPT and Codex text in the EU (October 2026)
• The Decoder — OpenAI will watermark ChatGPT text in the EU but makes it optional for API users worldwide
• EU AI Act — Article 50: Transparency obligations
• European Commission — Code of Practice on transparency of AI-generated content

Fact-checked against launch coverage of OpenAI’s textGrain announcement and the EU AI Act text as of October 8, 2026. OpenAI’s own announcement page blocks automated checks, so figures are cited via The Next Web and The Decoder. Rollout timing may change.