OpenAI textGrain watermarks ChatGPT text under EU AI Act

OpenAI has started adding an invisible watermark to text its models write. The system is called textGrain. From October 5, 2026, API customers anywhere can switch it on for their projects, and it stays off unless they do, Unite.AI reports. ChatGPT and Codex users in the European Union will get the watermark automatically in the coming weeks. The move answers an EU AI Act rule that makes AI-generated content identifiable by machines.

What OpenAI is switching on

The rollout comes in three parts, according to Unite.AI:

Who	What changes	When
API customers, worldwide	Can opt in to watermarking; it is off by default	From October 5, 2026
ChatGPT and Codex users in the EU	Watermarks added on every plan	In the coming weeks
Researchers and expert organizations	Can apply to use the detector, approved case by case	Applications opened October 5, 2026

Users outside the EU will not get watermarked ChatGPT or Codex text under this plan. Detector access runs under the EU's Code of Practice on transparency for AI-generated content, Unite.AI says. Crypto Briefing reports that OpenAI signed that code in June 2026.

How textGrain works

A text watermark hides a pattern in which words a model chooses. A reader cannot see it, but a detector can test a passage for it. Unite.AI says textGrain builds this statistical signal into the model's word selection.

How well detection works depends on the length of the text and how much it was changed. These are the figures Unite.AI reports, measured at a 1% false-positive rate:

Test	Passages detected
200-token passage	80%
400-token passage	95%
10% of words replaced	66%
25% of words replaced	17%

A token is a short chunk of text, often a whole word or part of one. A 1% false-positive rate means about one in 100 human-written passages would be wrongly flagged.

Edits are the weak point. "Text is easy to tweak, paraphrase or run through another tool, and each edit can weaken a watermark's signal," Crypto Briefing writes. OpenAI held back an earlier version for this reason. Crypto Briefing reports that a 2024 prototype was "about 99.9% effective," but OpenAI delayed its release over concerns about how robust it was.

What a watermark can and cannot prove

OpenAI lists clear limits, according to Unite.AI. A watermark cannot show that text is accurate, and it cannot establish who owns it. A missing watermark does not prove a human wrote the text. Detection is unreliable on short passages and on edited or translated text. It also fails on text from models the detector does not cover, including competitors' models.

Crypto Briefing raises a fairness concern too. People who write in a second language often use AI to polish their work, and watermarks could flag them.

Why now: Article 50 of the EU AI Act

Article 50 of the EU AI Act requires providers of generative AI to mark generated content in a machine-readable way. It has applied since August 2, 2026, and Crypto Briefing says it covers anyone serving EU users. City A.M. notes an exception: AI that only assists with editing, without substantially changing the original meaning, does not have to mark its output.

How rivals compare

OpenAI is not the first to watermark text. These reports describe the field:

Company	Approach	Source
Anthropic	Invisible watermarks on Claude text worldwide since August 2, 2026	City A.M., Crypto Briefing
Google	SynthID-Text watermarking in Gemini since 2024	Crypto Briefing
xAI	Refused to sign the EU transparency code, though still bound by the AI Act	City A.M.

City A.M. says Anthropic's marks survive copying and pasting, and that human-written text edited by Claude may carry them. Watermarks have drawn scrutiny before: SynthID's image watermark turned out to carry a 64-bit ID field, according to a research paper reported last month.

What this means for developers

If you build on OpenAI's API, decide whether to switch textGrain on. It is off by default. If your product serves EU users, ask your legal team whether Article 50 makes you responsible for marking AI output yourself. The API setting gives you one way to mark that output.

Do not treat the detector as a lie detector. Short or edited text slips through, and a missing watermark proves nothing. Any school, hiring or moderation tool that reads it as proof of human authorship will make mistakes.

If your team uses ChatGPT or Codex in the EU, expect text copied from them to carry the mark. Heavy editing weakens it, but light edits may not.

Watch who gets detector access. For now only approved researchers can test text. If OpenAI opens the detector wider, watermark checks could start appearing in publishing and education tools.