OpenAI will watermark ChatGPT and Codex text in the EU under the AI Act

OpenAI will add an invisible textGrain watermark to eligible EU ChatGPT and Codex text and lets API customers worldwide opt in for select models.

OpenAI said on October 5 that over the coming weeks it will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union, to meet the EU AI Act's machine-readable marking rules. If you ship ChatGPT- or Codex-written copy to or from EU teams, some of your deliverables will soon carry a mark that you can't see or check yourself.

The method, textGrain, adds a statistical signal to the model's word choices rather than hidden characters. TechCrunch reports that the mark travels with copied-and-pasted text. Outside ChatGPT and Codex, API customers globally can opt in for select models starting now. Watermarking stays off by default in the API. OpenAI hasn't said which outputs count as "eligible."

OpenAI's own tests show where the mark is weak. At a 1% target false-positive rate, it caught about 80% of 200-token passages and about 95% of 400-token ones for content such as psychology, and substantially fewer for math. Swapping 10% of words for synonyms cut detection from about 92% to 66%. Swapping 25% cut it to 17%. The detector isn't public. Only approved researchers and expert organizations can apply for access.

OpenAI says a detected mark doesn't measure human contribution, establish ownership or identify the user, and a missing mark doesn't prove human authorship. As Search Engine Journal points out, an agency with EU and non-EU writers on one plan may end up with marked copy from one office only. Contracts that treat detection as proof of authorship are asking the watermark a question it can't answer.

For comparison, Anthropic marks text from supported Claude models worldwide, according to the same report. Google has used SynthID-Text in Gemini since 2024, per a September arXiv paper.

Sources