OpenAI will watermark ChatGPT text written in the EU

Your ChatGPT drafts are about to carry a fingerprint in the EU.

OpenAI said on Monday that over the coming weeks it will add an invisible watermark to eligible ChatGPT and Codex text generated in the EU. Starting the same day, API customers anywhere can opt in to text watermarking for select models. In the API it stays off by default, and OpenAI says it will not become a global default at launch.

The watermark lives in the word choices

The technology is called textGrain. OpenAI says it adds a statistical signal through the model's word choices. That matters for anyone who copies and pastes. The signal sits in the wording itself, so a plain copy into another document carries it along.

A detector looks for that pattern. OpenAI says detector access will start out limited to approved researchers and expert organizations, because short passages and later edits make detection less reliable.

The company is blunt about the limits. The watermark can indicate that an OpenAI system generated or processed part of a text. It cannot tell who used the system, how much a person contributed, who owns the text or whether the content is accurate. And if no watermark is found, OpenAI says, that does not prove a person wrote it.

The numbers from OpenAI's own tests, as reported by ActuIA, show why. Detection ran at about 95% on a 400-token passage and about 80% at 200 tokens, at a 1% false positive rate. Replace a quarter of the words with synonyms and detection on 400 tokens fell to 17%.

17%.

Compliance first, detection second

This is the EU AI Act at work. ActuIA reports the move answers Article 50(2), the Act's rule on marking AI-generated content, which has applied since August 2, 2026.

Here's the thing: a watermark that a light edit defeats is not a cheating detector. It is a compliance signal. OpenAI is shipping it because the law asks for it, and it is telling you, in plain words, not to over-read it.

The real risk sits downstream. A client, a teacher or an HR team could treat a positive hit as proof of who wrote something. OpenAI says a hit cannot prove that.

What it means for you

If you use ChatGPT or Codex in the EU, assume your raw output will soon be detectable by the people who get detector access. Your edits will weaken the signal, and heavy edits may erase it.

If you build on the API, nothing changes unless you turn it on. If your product needs to show provenance, for a publisher or a regulator, you can now opt in on supported models.

If you run a team, write down your AI disclosure rule now, before someone else's detector writes it for you. A hit says an OpenAI system touched the text. It does not say what your person did with it.

What to watch

Watch the EU rollout dates for ChatGPT and Codex, and who gets detector access next. Then watch whether other labs selling into the EU ship their own text watermarks, and whether any of them work after an edit.

Write the policy this week.