OpenAI will watermark ChatGPT and Codex text in the EU
The AI Act is forcing the first large-scale provenance marking of chatbot text, with a detector researchers get first.
OpenAI said on October 5 that it will begin adding invisible watermarks to the text its models generate for users in the European Union, covering both ChatGPT and Codex. According to TechCrunch and an OpenAI post titled "Our approach to EU text provenance rules," this is the first time the company has watermarked chatbot text for a large group of users. A European law is the reason it is happening now, not a change of heart.
The method, which TechCrunch reported OpenAI calls textGrain, works by subtly shaping the model's word choices so the finished text carries a pattern. A reader cannot see it. Specialised detection systems can pick it out. TechCrunch reported OpenAI as saying the watermark does not identify the user who prompted the model, and that the company saw no meaningful change in model performance with the feature switched on.
The trigger is the EU AI Act, which took effect on August 2. TechCrunch reported that the watermarking is meant to comply with the law's transparency requirements, the part of the rulebook that pushes providers to make machine-generated content identifiable as such. For OpenAI, that turns a long-debated idea into a shipping feature.
The marking will not arrive everywhere at once. Per TechCrunch, it rolls out to eligible EU ChatGPT and Codex users over the coming weeks. Developers who build on OpenAI's API can already turn the feature on anywhere in the world, but for them it is off by default, so the choice sits with them. Many EU users will see no change for weeks yet.
OpenAI was candid about what the watermark cannot do. According to TechCrunch, replacing 10% of the words in a passage cut detection from 92% to 66%. Short passages, math answers and translated text are all harder to detect, and a missing watermark does not prove that a human wrote the text. "Editing can make the invisible marks harder to detect," OpenAI said, an admission that a determined user can wash the signal out.
The detector that reads these marks will not be open to everyone, at least at first. OpenAI's post says access starts with researchers, and TechCrunch reported that the company will initially give it only to approved researchers and expert organisations. That keeps the tool away from anyone who might use it to reverse-engineer and defeat the watermark.
OpenAI is not the first down this road. TechCrunch noted that Anthropic announced similar watermarking two months earlier. Together the two moves suggest invisible text provenance is becoming the industry's shared answer to the same regulatory pressure, not a one-company experiment. The parallel timing points to a European rulebook, rather than competition, setting the industry's pace on provenance.
Sources
- OpenAI will start watermarking ChatGPT’s text in the EU · TechCrunch, AI
- Our approach to EU text provenance rules · OpenAI News