OpenAI brings ChatGPT text watermarking to the EU and opens API opt-in
OpenAI will add an invisible watermark to qualifying ChatGPT and Codex text inside the European Union over the coming weeks, and from now API users worldwide can switch watermarking on for selected models. It stays off by default, and the EU rollout is the only place where the mark is applied automatically.
The method, called textGrain, adds a statistical signal to the model's word choices that a detector can look for. At a 1% target false positive rate, OpenAI's detector found the mark in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology, with much lower rates for subjects like maths where word choice is less flexible.
Editing weakens the signal. In 400-token English passages, swapping 10% of words for synonyms cut detection from roughly 92% to 66%, and swapping 25% cut it to 17%. The detector is not public at launch and is limited to approved researchers and expert organisations, who can ask for access case by case.
The timing follows Article 50 of the EU AI Act, whose transparency rules began applying on 2 August 2026. Systems already on the market have until 2 December to meet the marking and detection obligation. Anthropic marks text from supported Claude models worldwide because it has no durable way to scope watermarking by region, so a team split between Berlin and Toronto on one ChatGPT plan could soon find some of its copy carries the mark and some does not.
OpenAI has said a watermark does not measure how much a person contributed, and a missing one does not prove a person wrote the text. Any AI policy that treats a detection result as proof of authorship is asking the watermark a question it was not built to answer.
Reported by Search Engine Journal ↗
Topics: ChatGPT, Search policy, Measurement