OpenAI is rolling out text watermarking for some of its chat and coding outputs in the European Union. In plain terms, over the coming weeks, eligible ChatGPT and Codex text generated for EU users will carry an invisible marker designed to help machines recognize that the text came from an OpenAI model.
The rollout is being done in stages. Starting now, API customers around the world can opt in to text watermarking for select models, but the feature remains off by default. For consumer-style usage, the automatic watermark is coming to eligible ChatGPT and Codex text in the EU in the coming weeks. OpenAI is also accepting applications for access to its text watermark detector, with early access limited to approved researchers and expert organizations.
The technology behind it is called textGrain. Instead of attaching a visible label, it adds an invisible statistical signal in the model’s word choices. A separate detector can then look for that signal to assess whether a passage contains an OpenAI watermark. OpenAI says textGrain matched or outperformed other approaches it tested, including text SynthID.
That does not make it a reliable AI-or-not checker in everyday situations. Shorter passages are harder to classify, and detection rates are lower for content with less flexible wording, such as mathematics. Editing also weakens the signal. In OpenAI’s evaluation of 400-token passages, replacing about 10% of words with synonyms cut detection from roughly 92% to 66%; replacing 25% of words brought it down to 17%.
This is why OpenAI is keeping detector access narrow at first. The company is framing text watermarking as an early, imperfect transparency tool rather than a final answer. It can provide a useful provenance signal, especially under the EU’s new rules, but false positives and false negatives remain real possibilities.
Image and audio verification are not changing in the same way. OpenAI says its existing tools for images and audio, including openai.com/verify and the Content Provenance API, will remain publicly accessible for organizations checking whether those files were generated by its systems.