OpenAI said Monday it will begin embedding an invisible watermark in text produced by ChatGPT and Codex for users in the European Union, responding to AI Act transparency rules that took effect on August 2. The company says the mark, dubbed textGrain, subtly influences word choices rather than adding a visible label, allowing a detector with the relevant key to identify the pattern after text is copied and pasted.¹

The rollout, planned over the coming weeks, makes watermarking automatic for eligible EU users across paid and free plans. Outside the bloc, OpenAI is taking a far more cautious route: API developers can opt into the feature for selected models, but it remains off by default. That contrasts with Anthropic’s earlier decision to apply watermarking to Claude outputs worldwide.

The regulatory logic is straightforward. Brussels wants AI-generated material marked so that other systems can recognize it. The technical reality is less tidy. OpenAI’s detector will initially be restricted to approved researchers and expert organizations, a limitation the company says reflects the need to test reliability and responsible uses. “These limitations contribute to our decision to provide initial detector access only to approved researchers and expert organizations,” OpenAI said.²

Its own tests explain the caution. In the strongest case, textGrain detected marked text about 92% of the time. But replacing 10% of words with synonyms cut that figure to roughly 66%; more extensive revisions weakened it further. Short passages, translations and mathematical answers also pose tougher detection problems.

That leaves the watermark as evidence, not a verdict. OpenAI stresses that an absent signal “does not prove human authorship,” and that a detected signal cannot settle how much human judgment or creativity shaped the final passage.¹ The EU policy may force disclosure into the product, but editing remains the practical escape hatch.