OpenAI to Watermark ChatGPT and Codex Text in EU Under AI Act
OpenAI will watermark ChatGPT and Codex text in the EU under the AI Act, while tests show light editing can sharply weaken overall detection reliability.
Summary
On October 5, 2026, OpenAI said it would add invisible text watermarks to ChatGPT and Codex for eligible users on all plans in the European Union, complying with EU AI Act transparency rules effective August 2. The rollout begins over the coming weeks. API developers worldwide can enable watermarking today for select models, but it is off by default and will not become a global default at launch.
OpenAI’s textGrain method subtly changes word choices, creating a pattern that survives copying and can be detected from the text with a secret key. Its technical report, co-written with University of Pennsylvania and Yale researchers, says the watermark does not identify users or meaningfully affect model performance. Reliability remains limited: replacing 10% of words with synonyms reduced detection from about 92% to 66%, while short passages, math answers and translations are harder to identify. Detector access initially goes only to approved researchers and expert organizations. A missing watermark does not prove human authorship, and detection cannot measure how much human judgment, editing or creativity shaped a passage. Anthropic introduced worldwide watermarking for Claude two months earlier, prompting complaints from users who viewed Claude as a tool for their own instructions and decisions. By 2024, OpenAI had built but withheld an earlier watermark partly over fears users would move to rivals. Anthropic, Google, Meta, Microsoft and OpenAI have committed to the EU code of practice for AI-generated content.
Positives
- Eligible ChatGPT and Codex users on every EU plan will receive invisible watermarking over the coming weeks.
- API developers worldwide can enable watermarking today for select OpenAI models.
- OpenAI found no meaningful model performance change when textGrain was enabled.
- TextGrain’s pattern remains embedded when generated text is copied and pasted.
- Anthropic, Google, Meta, Microsoft and OpenAI have committed to the EU code for AI-generated content.
Risks & concerns
- Replacing 10% of words with synonyms cut detection from about 92% to 66% in OpenAI testing.
- Short passages, math answers and translated text are harder for textGrain to detect reliably.
- A missing watermark cannot establish human authorship because editing, length or another company’s AI can defeat detection.
- Initial detector access is restricted to approved researchers and expert organizations because of reliability and responsible-use concerns.
- OpenAI previously withheld watermarking partly because it feared users would switch to competitors without similar requirements.