Is There a ChatGPT Watermark? Only in the EU, for Now

Last updated October 6, 2026

Back in August I wrote about Claude watermarking everything it writes, and my question then was whether fixing a comma should count. On Monday OpenAI joined in, sort of. It published “Our approach to EU text provenance rules” and said ChatGPT text in Europe is about to carry an invisible ChatGPT watermark.

If you’re in the US, your ChatGPT answers aren’t marked. Claude’s are, though, and so are Gemini’s. So whether the thing you just pasted into an email carries a hidden signal depends on which chatbot wrote it and where you live.

Is ChatGPT text watermarked now?

In the EU, yes, starting over the next few weeks. OpenAI says the mark goes on “eligible ChatGPT and Codex text output in the European Union,” and TechCrunch reports it covers users on all plans, free included.

You can’t see it. textGrain nudges which words the model picks using a secret key, and over a few hundred words those nudges add up to a pattern a detector with the same key can spot. OpenAI put the math in a technical report written with researchers from Penn and Yale.

The reason is the EU AI Act. Article 50 says AI output has to be “marked in a machine-readable format and detectable as artificially generated,” and that rule took effect August 2. Systems that were already on the market get until December 2, which is why OpenAI can roll this out over weeks instead of flipping it on all at once.

Does the ChatGPT watermark affect you in the US?

Not right now. OpenAI told TechCrunch it is “not making text watermarking a global default at launch,” and it hasn’t given anyone outside the EU a date.

The one wrinkle is that developers who build on OpenAI’s API anywhere in the world could switch the watermark on for some models starting October 5. It’s off unless they choose it, but it means an app built on OpenAI could hand you marked text even if ChatGPT itself doesn’t.

Can a teacher or boss detect a ChatGPT watermark?

Not with OpenAI’s tool. The detector is going only to “approved researchers and expert organizations” for now, so there’s no public checker that you or your kid’s teacher can paste an essay into.

And even with access, it misses a lot. OpenAI’s own numbers show that on a test set of everyday-question answers it caught about 80% of 200-token passages (around 150 words) and about 95% at 400 tokens, while wrongly flagging human text 1% of the time. Math answers did worse, around 60%. In a separate test where it started at about 92%, swapping 10% of the words for synonyms dropped detection to 66%, and swapping a quarter of them dropped it to 17%. Translation hurts it too. ActuIA reports that OpenAI’s tests of English text translated into other EU languages caught it anywhere from 42.2% of the time in Romanian to 69.0% in Spanish, and that the EU rules don’t cover really short text under 200 tokens at all.

OpenAI is upfront about the other direction too. It says “the absence of a detected watermark does not prove human authorship,” and BleepingComputer reports that a hit doesn’t reveal who generated the text, their account, or the conversation behind it. I couldn’t find anything from OpenAI about letting EU users turn it off.

Claude and Gemini already watermark text

Both already do it, and Anthropic’s applies worldwide. Anthropic marks Claude’s text “wherever Claude is offered, worldwide,” across the Claude app, the API, Claude Code, Cowork, and Claude Tag. Google uses SynthID on text from the Gemini app and web.

Their detectors are locked down too. Anthropic’s is a private preview for regulators, media, researchers, and educational organizations. A school could apply, but a single teacher can’t, and Google’s SynthID Detector still runs off a waitlist for journalists and researchers. Each one only reads its own mark, because as Anthropic puts it, another company’s watermark “would have a different key.”

Not everyone took Claude’s version well. TechEconomy reported complaints on Reddit from people worried it could expose their AI use at work or school, and Decrypt found a GitHub tool that runs Claude’s text through a different model to strip the mark.

Where I land on it

I’m fine with this. Most people use AI at this point, so I don’t see the point of hiding it, and if you’re afraid of transparency then maybe AI isn’t the tool for you.

My problem was always how fine-grained it gets. I type a lot of my emails myself and hand them to Claude to clean up the grammar, and if that turns into “generated by Claude,” it’s like an editor lightly editing a book and then the book gets called the editor’s. Anthropic’s own write-up actually helps on this one. It says light proofreading leaves “very little (if anything) for the watermark to attach to,” which makes sense, since the mark lives in word choices and a grammar fix keeps most of yours. OpenAI, per The Next Web, says its watermark doesn’t measure human contribution either.

As an engineering manager, I don’t care whether someone on my team wrote code line by line or used AI, as long as it’s clean and it ships. Whether any of this matters for regular people comes down to how easy detection gets, and right now nobody outside a short list can even run the check.

FAQ

For now the ChatGPT watermark is a European thing, and the Claude and Gemini ones are everywhere. If OpenAI turns it on for US accounts, I’ll add it here.

Related reading: Claude watermarks everything it writes now | Is Claude’s writing getting worse? | How to turn off Gemini | New to AI? Start here