OpenAI to watermark ChatGPT and Codex text in the EU under AI Act; API opt-in available worldwide

OpenAI will begin adding an invisible watermark to text generated by ChatGPT and Codex in the European Union to comply with the EU AI Act, the company said Monday in a blog post. The EU AI Act's transparency rules, which took effect August 2, require AI companies to mark AI-generated content in a way other systems can identify.

Rollout details

• ChatGPT and Codex: The watermark will roll out "over the coming weeks" to eligible users on all plans, in the EU only.
• API: Developers anywhere in the world can enable watermarking on select models starting today. It is off by default. OpenAI has not specified which models are supported in the material we reviewed.
• Global default: OpenAI said it is not making text watermarking a global default at launch.

How textGrain works

OpenAI published a technical report alongside the announcement describing its method, called textGrain. It was co-written with researchers from the University of Pennsylvania and Yale. The watermark is not a visible symbol. It subtly shapes the model's word choices, using a secret key to sort next-word predictions. Across hundreds of these small nudges, a detector with access to the text and the key can identify the content as AI-generated.

Because the signal sits in the word choices themselves, it persists when text is copied and pasted. OpenAI says the watermark does not identify the user, and that it saw no meaningful change in model performance with watermarking enabled. Both are company claims; no independent evaluation has been published.

Limitations OpenAI disclosed

The watermark is not robust to editing. In one OpenAI test, replacing 10% of words with synonyms dropped detection from about 92% to 66%. The company also said short passages, math answers and translated text are harder to detect.

OpenAI cited these limits as the reason it is giving initial detector access only to approved researchers and expert organizations. The company also cautioned that a missing watermark "does not prove human authorship," since text may be too short, too heavily edited, or generated by another company's AI. A watermark can indicate that an OpenAI system generated or processed part of a passage, OpenAI said, but not how much human judgment or editing went into it.

Industry context

The announcement comes two months after Anthropic said it would watermark text generated by Claude, a policy Anthropic is applying worldwide. That decision drew backlash from some Claude users, who argued they had supplied the instructions, context and decisions while Claude was only the tool.

OpenAI had built a text watermark before but held off on releasing it. The Wall Street Journal reported in 2024 that one reason was concern that users would move to rivals that did not watermark. Anthropic, Google, Meta, Microsoft and OpenAI are among the companies that have committed to the EU's code of practice on AI-generated content.

What this means

This is a compliance-driven rollout, and the scoping shows it. OpenAI is limiting consumer-facing watermarking to the EU, where the law requires it, and leaving the API as an opt-in. That avoids the competitive risk it reportedly worried about in 2024. Anthropic's global approach makes OpenAI's regional one a notable divergence.

The technical picture is more limited than the regulatory one. A drop from about 92% to 66% detection after a 10% synonym swap means textGrain can flag unedited output but is weak evidence against anyone who edits. Restricting the detector to approved researchers also means educators, publishers and platforms cannot yet verify text themselves. For developers, the practical question is which API models support the feature and whether downstream pipelines that rewrite, translate or summarize output will strip the signal. For now, the watermark is an AI Act compliance mechanism, not a reliable authorship test.