ChatGPT text in Europe is about to carry an invisible signature

OpenAI said on Sunday that text produced by ChatGPT and Codex inside the European Union will start carrying a hidden marker. Nothing about the writing looks different. The system, called textGrain, nudges which word the model picks at each step so the finished text holds a statistical pattern that a detector with the right key can spot.

The reason is Article 50 of the EU AI Act, which says machine-generated content has to be identifiable in a machine-readable way, with the obligation applying from 2 December. The rollout reaches eligible users on all plans in the EU over the coming weeks. Developers outside Europe can switch the same thing on through the API if they want it, off by default. The detector itself is not being handed out publicly yet, only to approved researchers and organisations.

OpenAI is unusually candid about how fragile this is. Swap one word in ten for a synonym and detection falls from about 92 percent to 66. Swap one in four and it drops to 17. Short passages are harder, maths is close to hopeless, and translating or heavily editing a piece can wash the mark out altogether. The watermark also says nothing about who wrote the text, which account produced it or what the conversation was about.

In practice this is something to understand rather than act on. A teacher, an employer or a newspaper cannot wave a wand over a document and get an answer, and anyone being told otherwise is being sold something. If you use the ChatGPT desktop app on Windows, Tendvane's app update check is what quietly keeps it on the current version as changes like this ship.