OpenAI will watermark ChatGPT text in the EU, and API watermarks are opt-in. Your content pipeline needs a policy

AI text is starting to carry a hidden label, and the rules differ by vendor, region and product.

What happened

OpenAI said on Monday it will add an invisible watermark to text generated by ChatGPT and Codex in the European Union, to comply with the EU AI Act transparency rules that took effect on August 2. It rolls out over the coming weeks to eligible users on all plans, in the EU only.

Developers using the OpenAI API anywhere in the world can turn watermarking on for select models. It is off by default.

The method, called textGrain, subtly shapes the model's word choices. Readers cannot see it, but a detector holding a secret key can. Because it lives in the words, it travels with copied text. OpenAI says it does not identify the user.

Detection is far from perfect. In OpenAI's tests, replacing 10% of words with synonyms dropped detection on 400-token passages from about 92% to 66%, and The Decoder reports that replacing a quarter of the words dropped it to 17%. Short passages, math and translated text are harder to detect. For now, only approved researchers and expert organizations get detector access, and OpenAI says a missing watermark does not prove a human wrote the text.

TechCrunch notes Anthropic announced two months ago that it would watermark Claude text worldwide.

My take

If you build content pipelines for clients, think blog drafts, outbound email sequences, product descriptions or CRM follow-ups, this is now a configuration decision, not trivia.

Three things I would do this week:

1. Map which provider and which product each workflow uses. ChatGPT in the EU, the API with watermarking on or off, and Claude all behave differently now.
2. Write the choice down per client. If an EU client asks how their AI content is labelled, "it depends on the node" is not an answer. Document the setting and the reason.
3. Do not build detection into your process. OpenAI itself says results can miss marks or flag unmarked text. A detector score is not proof of anything, so never use one to judge a writer or a supplier.

The bigger point: watermarks say a model was involved, not how much human judgment went in. The value you sell clients is the review, the facts and the voice, so make that step visible in the workflow instead of hiding the AI part.