Skip to content

AI

OpenAI will secretly sign your ChatGPT text, EU users first

OpenAI is rolling out textGrain, an invisible watermark for ChatGPT and Codex output in the EU. It is built for the AI Act, and OpenAI itself admits detection is not a sure thing.

Cover photo: Jernej Furman from Slovenia / CC BY 2.0

TL;DR

  1. OpenAI adds an invisible textGrain watermark to ChatGPT and Codex text and code in the EU.
  2. It exists to satisfy the AI Act's transparency rules; Anthropic made a similar move earlier.
  3. Short text lands around 80 percent detection, and editing the text makes it worse.

What happened

OpenAI says text and code its tools produce inside the European Union will soon carry a watermark nobody can see. The system is called textGrain.

It nudges the statistical pattern of the model's word choices, and a detector can later read that pattern and flag the text as machine-made.

Over the next few weeks it reaches eligible ChatGPT and Codex accounts in the EU, free and paid alike.

OpenAI says it is not making watermarking a global default at launch. The watermark will not be on by default either, except for eligible ChatGPT and Codex text in the region.

The motive is compliance. OpenAI is meeting part of the EU's AI Act transparency rules, following a similar move by Anthropic earlier in 2026.

OpenAI claims textGrain matched or exceeded other approaches, including Google DeepMind's SynthID for text. It also published benchmark scores showing similar performance for watermarked and unwatermarked output.

Then comes the fine print, and it is generous. OpenAI warns that strong results under ideal conditions do not guarantee reliable detection in everyday use.

Shorter text is harder to catch, at around an 80 percent success rate. Mathematics-style content is harder too, and editing the text also lowers the odds. OpenAI adds that a watermark does not verify accuracy, establish ownership, measure human contribution, or prove human authorship.

OpenAI also plans to release the technology as open source so others can tweak it.

Why it matters

This is the first big look at how a frontier lab turns AI Act language into an actual product feature. When a leading provider ships it, the rest of the industry gets a template.

It is also a reality check. A watermark that fades when someone rewrites a paragraph is a speed bump, not a lie detector. Anyone hoping to catch AI text with this should read OpenAI's own caveats first.

The reaction was not uniformly warm either. Not everyone was happy when Anthropic announced its watermarking.

What it means for Europe

Europe is where this lands first, and that is the point. The AI Act is now changing product design, not just legal memos, and EU users get the feature before anyone else does.

For businesses and developers in the EU, that means text and code from ChatGPT and Codex may carry a hidden signal that detection tools can find. Teams that publish AI-assisted content should know that before a compliance question arrives.

OpenAI is keeping this EU-only at launch, so the rest of the world gets unmarked text for now. Brussels writes the rules, and the models adjust their handwriting.

What to do with it

  • Check your plan.If you use ChatGPT or Codex in the EU, watch for the rollout over the coming weeks and see whether your output is eligible.
  • Do not treat it as proof.Detection is imperfect, so never use a missing or present watermark as the only evidence of who wrote something.
  • Write your disclosure policy.If your team publishes AI-assisted text or code, decide now how you label it, before a detector does it for you.
  • Watch the open-source release.OpenAI plans to publish the technology, so developers can follow it and see how it works.

Ciao, sapiens. The machines now sign their homework. Try not to forge the signature.