ChatGPT to Add an Invisible Text Watermark in the EU

OpenAI said on October 5, 2026 that text generated by ChatGPT and Codex in the European Union will carry an invisible watermark, called textGrain, over the coming weeks. It is not a global default. For API customers it is opt-in worldwide from the same day. It matters to anyone who publishes text written with AI.

What we know

  • Who and when: OpenAI published it on October 5, 2026 in its post "Our approach to EU text provenance rules". It says the move responds to the EU AI Act, which requires generated text to be identifiable in a machine-readable way.
  • ChatGPT and Codex: the watermark arrives "over the coming weeks" on eligible ChatGPT and Codex text, on all plans and in the EU only. The post says it will not be a global default at launch.
  • API: from October 5, API customers anywhere can opt in for select models. It stays off by default, according to OpenAI.
  • How it works: textGrain adds an invisible statistical signal to the model's word choices. According to OpenAI's help page, it adds no hidden characters, invisible spaces or unusual punctuation.
  • What is left out: the EU code of practice does not require watermarks in outputs shorter than 200 tokens (about 150 words in English) or in code snippets, the help page says.
  • The detector is not public: OpenAI is taking applications only from approved researchers and expert organizations, case by case.
  • How well it detects: at a 1% false positive rate, the detector found the watermark in about 80% of 200-token passages and about 95% of 400-token passages, for content such as psychology. For mathematics it was substantially lower, the post says.
  • Editing weakens it: in 400-token passages, replacing 10% of the words with synonyms cut detection from about 92% to 66%. Replacing 25% cut it to 17%.
  • By language: in a test across the 24 official EU languages, Spanish had the highest detection rate, 69.0%, and Romanian the lowest, 42.2%, according to the help page.
  • The rule: the European Commission says the transparency obligations in Article 50 of the AI Act apply from August 2, 2026.

The detection figures are OpenAI's own measurements of its own technology, not a third party's.

What changes and what doesn't

What changes: ChatGPT text generated in the EU will carry a signal that a detector holding the key can look for.

What you see does not change: the text reads the same. OpenAI says its tests showed no meaningful quality differences with and without the watermark.

The watermark says little, by OpenAI's own account. It does not measure how much a person contributed. It does not establish who owns the text or who is responsible for it. It does not identify the user or the conversation. It does not verify that the text is accurate. And a missing watermark does not prove a person wrote it: the text may be short, edited or translated, or come from another company's tool.

Neither OpenAI's post nor its help page mentions search rankings. On what has been published, there is no basis to claim the watermark changes how a page appears in Google.

How to tell if it affects you

Five checks if you write copy for your site with ChatGPT, or if your site uses the OpenAI API:

  1. Check where ChatGPT is used from. If you and your team work outside the EU, OpenAI says the watermark is not on by default. If someone writes from an EU country, their eligible text will carry it once the change arrives.
  2. If your site uses the OpenAI API (a support chat, a product description generator), the setting is under Organization settings → Data controls → Text provenance, or Project Settings → Text provenance, according to the help page. It is off by default: decide with whoever runs the technical side.
  3. If you sell to customers in the EU, ask your legal adviser which notices apply to you. OpenAI warns that the watermark does not replace visible labels that may be required.
  4. Don't treat an "AI detector" as proof. OpenAI's is not open to the public. Third-party ones typically classify text by its patterns rather than reading an embedded signal, OpenAI explains. If you commission copy, ask for the source of every fact rather than a "human" percentage.
  5. Keep a record of how each text was made: which tool, what date, who reviewed it. If a client or a platform asks, the answer will be in your records, not in a watermark.

This article is not legal advice: what the EU AI Act requires of each business depends on its case.

Related: Google: Fact-Check AI Content by Hand Before Publishing

Sources

Updates: the date the watermark reaches ChatGPT in the EU, and any change to detector access, will be added here.