OpenAI will watermark ChatGPT and Codex text in the EU

OpenAI says it will add an invisible watermark to eligible text generated by ChatGPT and Codex in the European Union, responding to the EU AI Act’s requirement that generated text be identifiable in a machine-readable way.

What changes #

For API customers, watermarking is available as an opt-in feature worldwide for selected models. OpenAI says it is off by default in the API and is not becoming a global default for ChatGPT or Codex at launch.

The company’s method, called textGrain, adds a statistical signal by nudging a model’s word choices. A detector can then look for that signal in a passage. OpenAI is also opening applications for detector access, initially limited to approved researchers and expert organizations rather than the public.

Why it matters #

The rollout makes text provenance less theoretical for everyday AI products: European users may soon generate text that carries a hidden platform-level signal, while developers can decide whether watermarking fits their own disclosure obligations.

OpenAI is also trying to narrow expectations. It says a watermark does not identify the user, prove authorship, establish responsibility, or verify whether a passage is true. Its own tests show the signal can weaken quickly: in one evaluation, replacing 10% of words with synonyms reduced detection from about 92% to 66%, and replacing 25% reduced it to 17%.

That caveat is the point. Watermarking is becoming part of AI compliance, but it is still a probabilistic clue, not a courtroom-grade proof system.