The European Union’s AI Act transparency rules took effect on August 2, requiring providers to make AI-generated content identifiable by other systems. OpenAI is now using that deadline to test a deliberately limited answer: invisible text watermarking, rather than a global switch.¹

Over the coming weeks, eligible ChatGPT and Codex users across EU plans will receive watermarked output. The company is not making the feature a worldwide default at launch, saying the regional rollout offers space to learn from real-world use and feedback.² API customers elsewhere can opt in for select models, while approved researchers and expert organisations may apply for detector access.²

The system, called textGrain, does not stamp a visible label on an answer. Instead, it subtly steers word choices, creating a machine-readable pattern that remains when text is copied and pasted. OpenAI says benchmark results showed no meaningful performance gap between watermarked and unwatermarked text, and that textGrain “matched or exceeded” approaches including Google DeepMind’s SynthID for text.¹

But OpenAI’s caveats are as consequential as its launch. Its tests found that replacing 10% of words with synonyms cut detection from roughly 92% to 66%; short passages, mathematical answers and translated text also pose tougher cases.¹ That is why the company says textGrain “does not guarantee reliable detection,” and why detector access is initially restricted.²

The company’s central warning is blunt: a watermark can suggest that an OpenAI system generated or processed part of a passage, “but not how much human judgment, editing, or creativity went into it.”¹ A missing mark, it adds, does not prove human authorship. The experiment may improve disclosure, but it leaves the harder argument—who really made the work—unresolved.²