The invisible watermark shapes word choices, and API developers worldwide can enable it for select models now.

OpenAI will add an invisible watermark to text generated by ChatGPT and Codex in the European Union, the company said Monday. The move complies with the EU AI Act’s transparency rules, which took effect on August 2 and require AI companies to mark generated content so other systems can identify it. The watermark reaches eligible users on all plans over the coming weeks, but only inside the EU.

OpenAI published a technical report for its method, called textGrain, written with researchers from the University of Pennsylvania and Yale. The system uses a secret key to sort next-word predictions, and a detector spots AI-generated content from the text plus that key. Anthropic rolled out text watermarking for Claude worldwide two months earlier, a move that drew pushback from some users.

Developers using the OpenAI API anywhere can switch on textGrain for select models starting today, though OpenAI leaves it off by default and does not make it a global standard. The watermark travels with copied text because it lives in the word choices themselves. Short passages, math answers, and translated text resist detection, and OpenAI limits detector access to approved researchers and expert organizations.

OpenAI warns that a missing watermark does not prove human authorship, since text can be too short or heavily edited, or generated by a rival system. Editing strips the mark: swapping 10 percent of words for synonyms dropped detection from roughly 92 percent to 66 percent in company tests. The European code of practice on AI-generated content counts Anthropic, Google, Meta, Microsoft, and OpenAI among its signatories.

What matters

• OpenAI began rolling out an invisible watermark for ChatGPT and Codex text inside the EU.
• API developers can enable textGrain on select models today, but the feature stays off by default.
• Editing weakens detection, since swapping 10% of words for synonyms cut accuracy from 92% to 66%.

Why it matters

Editing weakens detection, since swapping 10% of words for synonyms cut accuracy from 92% to 66%.

This GenAI News article was prepared in original wording using reporting and materials published by TechCrunch AI. Source reference: https://techcrunch.com/2026/10/05/openai-will-start-watermarking-chatgpts-text-in-the-eu/.

Drafted by the GenAI News review pipeline.