OpenAI will hide watermarks in EU ChatGPT text, but edits can erase them

OpenAI is starting to put an invisible watermark in the text ChatGPT and Codex write for users in the European Union. If you're in the EU, your chatbot's answers will soon carry a signal you can't see, and for now almost nobody will have a tool to check for it.

The system is called textGrain. OpenAI says it nudges the model's word choices into a statistical pattern that a detector can pick up later. When several words fit equally well, a secret key decides which one gets used. No hidden characters are added. The watermark switches on automatically for ChatGPT and Codex in the EU over the coming weeks. API developers anywhere can opt in for select models, but it stays off by default there. OpenAI says textGrain "matched or exceeded" Google's SynthID in its own tests.

Light editing can wash it out

OpenAI is unusually open about the limits. According to figures OpenAI published with its announcement, detection works about 80% of the time on a passage of roughly 200 tokens (a few short paragraphs) and about 95% at 400 tokens, with a 1% false positive rate. Those numbers come from OpenAI's own tests on ideal text, not from independent checks. Replace 10% of the words with synonyms and detection drops to 66%. Replace a quarter and it falls to 17%. For now the detector is closed. Only approved researchers and expert organizations can apply, and OpenAI says it doesn't identify users or reveal prompts. OpenAI says it will eventually release the technology as open source, but it hasn't given a date.

Google, Anthropic and Meta face the same rule

Article 50 of the EU AI Act has applied since August 2, 2026. It requires makers of generative AI to mark their output in a machine-readable way that can be detected. Tools that only help with standard editing are exempt. Under the AI Act's transitional rules, generative AI systems that were already on the market before August 2 have until December 2, 2026 to add that marking. Engadget reports that this gives major providers, including Google, Anthropic and Meta, until then to comply. Anthropic went first. In August it said Claude models released after August 2 would carry a text watermark that stays with the text when it's copied and pasted. Google already watermarks text from the Gemini app with SynthID. Anthropic, Google, Meta, Microsoft and OpenAI are all on the Commission's list of providers that signed Section 1 of the Code of Practice on Transparency of AI-generated Content, a voluntary code published in June 2026.

That leaves an odd result. The same prompt can give you marked text in Paris and unmarked text in Toronto, because OpenAI's automatic ChatGPT rollout covers only the EU. And with detection that drops so quickly after edits, a missing watermark tells you very little about who wrote something.

The next things to watch are the ChatGPT switch-on over the coming weeks and the December 2 deadline for systems already on the market. After that, watch who gets into OpenAI's detector program, which opened to researcher applications on October 5, and whether OpenAI puts a date on its open source release.