Does ChatGPT watermark text? What the EU rollout means for your content

TL;DR

• Yes, in the EU. OpenAI is adding an invisible watermark to eligible ChatGPT and Codex text for EU users over the coming weeks, on every plan.
• Outside the EU, ChatGPT text isn't watermarked by default. Developers using the API anywhere can switch it on for select models, but it stays off unless they do.
• The watermark lives in word choice, so it survives copy and paste. Swapping 10% of the words for synonyms cut detection from about 92% to 66% in OpenAI's tests.
• Only approved researchers and expert organisations can use OpenAI's detector for now, and a missing watermark does not prove a person wrote the text.
• For marketers, what the text says carries more weight than how it was made. A draft built on your own material and properly edited sits comfortably with Google and the EU AI Act, no scrubbing required.

On 5 October 2026, OpenAI confirmed that ChatGPT will start watermarking its words. Two months earlier Anthropic did the same for Claude, worldwide. If you draft with AI, there's now a decent chance your copy carries a signature you can't see.

Here's what the ChatGPT watermark is, who it applies to, how well it survives editing, and what a small content team should do about it.

Does ChatGPT watermark text?

Yes, in the European Union. OpenAI is adding an invisible watermark to eligible ChatGPT and Codex output for EU users over the coming weeks. The watermark covers every plan and only applies in the EU, and it isn't a global default at launch.

Developers using OpenAI's API anywhere in the world can turn the watermark on for select models. It's off by default.

The trigger is the EU AI Act. Its transparency rules apply from 2 August 2026 and require providers of generative AI to mark output in a machine-readable way, so it can be detected as AI-generated. Systems already on the market before that date have until 2 December 2026 to meet the marking requirement.

How the ChatGPT watermark works

The watermark is a pattern in the words themselves, with no visible symbol and no tag in the file. Each time the model picks the next word, a secret key nudges it towards certain options among the plausible ones. A single nudge means nothing. Hundreds of them add up to a statistical fingerprint that a detector can spot using only the text and the key.

The method is called textGrain, built with researchers from the University of Pennsylvania and Yale. It works much like Google's SynthID, which already watermarks text from the Gemini app, and like the Claude watermark.

That design has two practical effects. Because the pattern lives in the wording, it travels with copy and paste into Google Docs, Word, your CMS, anywhere. And metadata strippers do nothing to it, because there's no metadata involved.

The watermark doesn't identify the user, and OpenAI saw no meaningful drop in model performance with it switched on.

What about the odd spaces in copied ChatGPT text?

Developers noticed unusual Unicode space characters in pasted ChatGPT text well before this rollout, and OpenAI put it down to a model quirk. That's a separate thing from the new statistical watermark. Copying out of any chat window, PDF, or Google Doc can pick up stray zero-width and non-breaking spaces too, and our free AI slop cleaner strips them in your browser. Just know what it can't do: cleaning up strange characters won't touch a pattern built into word choice.

Can you remove a ChatGPT watermark?

Editing weakens it a lot. Replacing 10% of the words with synonyms dropped detection from about 92% to 66% in OpenAI's tests. Short passages, maths answers, and translated text are also harder to detect.

That's part of why OpenAI is limiting its detector to approved researchers and expert organisations for now. A missing watermark doesn't prove human authorship either. The text could be too short, too heavily edited, or written by another company's model.

So a determined person can blur it. Paraphrasing through another model, or rewriting by hand, break the pattern. Text watermarks will always be easy to strip for exactly that reason.

Think about what that route gets a marketer, though. Run ChatGPT copy through a paraphraser and you end up with text that's still generic, now slightly clumsier, and still reads like every other result on page one. You've hidden where it came from and left everything else about it untouched.

Does a watermark hurt SEO?

There's no sign Google ranks watermarked text lower. Appropriate use of AI isn't against Google's guidelines. The line sits at content made mainly to manipulate rankings, and AI output should be fact-checked before it goes live.

Google runs its own text watermark through SynthID, and nothing in its search guidance treats watermarked pages differently. The ranking risk sits where it always has, with thin, interchangeable content published at volume, which is what makes AI content hurt SEO.

What the EU AI Act asks of you as a publisher

Marking the output is OpenAI's job as a provider. Publishers carry a separate set of obligations, and for typical marketing content they're lighter.

Under Article 50(4), anyone publishing AI-generated text with the purpose of informing the public on matters of public interest must disclose it, unless the text went through human review and a person or company holds editorial responsibility for it. The review has to be substantive. A glance and a thumbs up won't count.

A product blog or a LinkedIn update usually isn't public-interest publishing in the news sense, but this isn't legal advice, so check your own situation. The safest habit is the same either way: a real person reviews, edits, and signs off everything that goes live.

The draft Code of Practice also proposes separating fully AI-generated content from AI-assisted content, with different disclosure expectations for each. Work that starts from your own thinking and gets properly edited sits much more naturally in the second group.

What small content teams should do now

• Feed the model your material. Use ChatGPT to research, structure, and draft around things only you have, like customer calls, sales objections, your own numbers, or the opinion you'd defend over a pint. A non-commodity content strategy starts with collecting it.
• Write your voice down once. A knowledge base with tone rules, banned phrases, and examples, read on every run, does more for originality than any paraphraser. That's how you maintain brand voice with AI at volume.
• Edit for substance. Adding your own examples, cutting the generic sections, and rewriting the opening changes far more than 10% of the words. The watermark weakens as a side effect of the piece getting better.
• Keep a record of review. Note who edited and approved each piece, so you have an answer if anyone asks about editorial responsibility.
• Stop worrying about detectors. OpenAI's isn't public, and a negative result proves nothing anyway.

The ChatGPT watermark changes less for good content teams than the headlines suggest. If a draft only works as long as nobody knows a model wrote it, the problem was there before the watermark arrived. If it's built on your material and properly edited, there's nothing to hide. Start by preventing AI slop at the source.

Frequently asked questions

Does ChatGPT watermark text?

Yes, for users in the EU, rolling out from October 2026 across every plan. Outside the EU it isn't on by default, though API developers can switch it on for select models.

Can you see the ChatGPT watermark?

No. It's a statistical pattern in the model's word choices, with no visible mark and no file metadata.

Can I check whether my text has a ChatGPT watermark?

Not yet, unless you're an approved researcher. At launch OpenAI's detector is limited to approved researchers and expert organisations.

Does Claude watermark text too?

Yes. Anthropic announced text watermarking for Claude about two months before OpenAI's EU rollout, and applies it worldwide.

Will Google penalise watermarked content?

There's no sign of it, and appropriate use of AI is within Google's guidelines. Helpfulness, accuracy, and originality still decide rankings.