ChatGPT Text Will Carry an Invisible Watermark in the EU

OpenAI said on Monday that it will hide a watermark in text and code produced by ChatGPT and Codex in the European Union. You won’t see anything different on screen, but a specialized detector can pick up the pattern later. The change rolls out over the coming weeks and is aimed at meeting the EU’s AI transparency rules.

How the watermark works

The technology is called textGrain. Instead of stamping a label on the text, it nudges the words the model picks, leaving a statistical pattern that readers can’t notice but a detector can flag. Think of it as a signal baked into word choice, not something printed on the page.

OpenAI says model performance doesn’t change in any meaningful way with the feature on. The company also plans to release textGrain as open source, which would let outside developers study and test the method.

In the EU, the watermark will be mandatory for ChatGPT users on every plan. API customers, meaning the companies and developers who build their own products on OpenAI’s models, get a choice. According to The Decoder, API customers worldwide can opt out, unlike with Anthropic. Xataka notes that models Anthropic has released in the EU since August 2 already ship with an invisible mark in their text.

What the detector can and can’t do

The numbers OpenAI shared describe a useful tool that is far from foolproof:

• Early tests showed roughly 92% detection, per TechCrunch.
• Swapping just 10% of the words for synonyms dropped that to 66%.
• The Decoder cites about 95% accuracy on 400-token passages about psychology (tokens are the word fragments a model processes) and around 80% at 200 tokens.
• Math answers do worse, at roughly 60% for 400 tokens.
• When a quarter of the words are replaced, detection falls below 20%, even in longer text.

Short passages, math and translated text are also harder to catch. OpenAI itself says a missing watermark does not prove human authorship. So the detector is a clue, not a verdict. Heavily edited or translated text can slip through, and that tells you little about who wrote it.

Who gets the detector

For now, you won’t be able to paste a paragraph into a website and get an answer. Access to the detector will initially be limited to approved researchers and specialist organizations, with requests reviewed case by case through an application form under the EU’s code of practice.

Why now

The backdrop is the EU’s AI Act, which requires AI-generated content to be machine-identifiable. Engadget reports that OpenAI’s move follows a similar announcement from Anthropic. The sources disagree on the exact compliance deadline, so we’re leaving the date out.

The announcement is about EU users. People using ChatGPT elsewhere don’t get the mandatory mark, though API features offer it as an option. It’s worth watching whether OpenAI extends the practice to other regions, since the technology doesn’t depend on geography once it’s built.

What this means in practice

If you study or work in Europe, using ChatGPT will now leave a trace. Editing the output heavily cuts the odds of detection, though, which shows the limit of the approach: it helps spot raw machine text, not police every sentence written with AI help. To see why these models sometimes make things up, our explainer on what a language model is covers the basics. And for another recent product change, see how ads are coming to the image generator.