OpenAI is rolling out an invisible watermark called textGrain for ChatGPT and Codex output that will be active for users in the European Union. The company says the feature will be switched on over the next few weeks to meet the EU AI Act’s requirement that generated text be labelable in a machine‑readable way. For developers who use OpenAI’s API, the watermark can be enabled or disabled at will, a choice that differs from Anthropic’s approach, which forces a watermark on all Claude outputs worldwide.
How the watermark works and its limits
textGrain inserts a statistical pattern into the model’s word choices, similar to Google’s SynthID used by Anthropic’s Claude. OpenAI’s internal tests say the detector finds the pattern in about 95 % of 400‑token (roughly 300‑word) psychology passages, but only about 60 % of math passages of the same length. Changing even a small part of the text reduces detection: swapping 10 % of words with synonyms drops the hit rate from roughly 92 % to 66 %, and replacing a quarter of the words cuts it to about 17 %.
Access to the detector
For now, OpenAI will give detector access only to selected researchers and specialist organisations through an application process. The tool will simply report whether an OpenAI watermark is present; it will not reveal who wrote the text or any prompts used.
Why it matters
EU users will see a hidden marker on ChatGPT output that regulators can read, helping meet new labeling rules. Developers can choose to hide the marker, which may affect how easy it is for third parties to spot AI‑generated text. The low detection after editing means users could still remove the watermark by modestly rephrasing content.
