TechCrunch reports that OpenAI will watermark text generated by ChatGPT and Codex for EU users, in line with EU AI Act transparency rules. API developers can opt in on some models.
What was announced
On October 5, TechCrunch reported that OpenAI will watermark text generated by ChatGPT and Codex using a technique called 'textGrain'. It introduces a slight bias in word choice that readers cannot see but that can be detected afterward.
According to the report, the rollout for EU users will happen gradually over the coming weeks. API developers can reportedly enable it on some models, but it is off by default.
Background and limits
The report says the move is tied to the transparency requirements of the EU AI Act, which call for AI-generated content to be identifiable. It also notes that Anthropic announced a watermark for Claude in August.
The approach has clear limits. The article says that rewriting about 10% of the words drops detection from roughly 92% to 66%. Short texts, math answers, and translations are harder to detect, and a missing watermark does not prove a human wrote the text. For now, detection tools are limited to approved researchers and specialist organizations.
TAKEAWAY
What this means for your project
If you commission or run a website that uses AI-drafted copy, do not treat a watermark, or its absence, as proof of quality or accuracy. Have a named person check prices, hours, and claims against your own records before publishing, and write down your AI policy: which tools, for what, and who signs off.