OpenAI watermarks ChatGPT and Codex text, starting with Europe

A mark nobody can see

On Monday 5 October 2026, OpenAI started rolling out textGrain, an “invisible, machine-readable” watermark woven into text produced by ChatGPT and Codex. The announcement was reported by The Verge. To the naked eye, nothing changes: no footer, no disclaimer, no suspicious characters lurking between your words. The mark is written for software to read, not for you.

OpenAI isn’t publishing the technical details of how the marking works. It does claim that textGrain “matched or exceeded” other approaches, such as Google DeepMind’s SynthID for text. SynthID also happens to be the basis of the watermark Anthropic announced in August, likewise to comply with the EU AI Act, and likewise greeted with a fair amount of scepticism. OpenAI has published benchmark results in which watermarked text performs on par with unmarked text: in other words, the model supposedly doesn’t write any worse just because it’s signing its homework.

Europe first, everyone else on request

The rollout will happen “in the coming weeks” for eligible ChatGPT and Codex users in the European Union, whatever their subscription plan. OpenAI says the watermark won’t become the “global default” for now: the regional approach is meant to let it learn from real-world use and feedback.

Everywhere else, it works the other way round: the option is opt-in. Since 5 October, API customers worldwide can choose to receive watermarked text for certain models, depending on their own transparency obligations. OpenAI is also working with its cloud partners to offer the same option on models served through their platforms, again in the coming weeks. The upshot is slightly surreal: the exact same text is either marked or not, depending on which country the person asking is in and which door they came through.

A detector for specialists only

If you want to check for the mark, you’ll need to show your credentials at the door. The detector is reserved for researchers and specialist organisations, who have been able to apply for access since 5 October. Access will be granted case by case, in line with the code of practice that accompanies the AI Act. The tool only answers one question, whether or not it found an OpenAI watermark, without identifying the user or revealing their prompts or conversations.

There’s no public version at launch, and OpenAI explains why: detection can miss watermarks and can also produce false positives. The company even admits that textGrain “does not guarantee reliable detection”. For a tool whose entire job is detecting things, that’s a refreshingly honest piece of small print.

What the mark doesn’t prove

If you publish generated text, the watermark’s reach is deliberately narrow. According to OpenAI, it doesn’t verify that the content is accurate, doesn’t say who owns the text, doesn’t measure how much human work went into it, and doesn’t prove that a human wrote it. The watermark says an OpenAI model produced the text, and nothing more: not whether it’s right, not who checked it.

In practice, if you write in Europe with ChatGPT or Codex, your text will soon carry this mark without you lifting a finger. Only a handful of approved labs will be able to read it, and even they get no guarantee it’s reliable. Think of it as a hand stamp at a nightclub: you can’t see it, the bouncer keeps the UV torch in a locked drawer, and every so often he waves the wrong person in.