ChatGPT in Europe Is Getting an Invisible Watermark. Here's What It Means

OpenAI will add a hidden watermark to ChatGPT and Codex writing in the European Union in the coming weeks. OpenAI The system is owned by OpenAI and called textGrain. Engadget It applies to normal writing and to computer code.
The move is linked to new EU transparency rules. Article 50 of the EU AI Act says makers of AI systems must mark AI-written text so computers can spot it. People will not see the mark. The rule starts on 2 August 2026. European Commission Services that already exist have until 2 December to comply. OpenAI says it supports the EU Code of Practice on AI content transparency, which helps with marking and labelling duties. Its newsroom lists the update as Our approach to EU text provenance rules, dated 5 October 2026.
textGrain works by quietly shaping word choices to leave a hidden pattern, like a faint stamp that only shows under special light. A detector tool can look for that pattern to check for the watermark. The detector does not name the user and does not show prompts or chats. At first, only approved researchers and expert groups can use the detector.
OpenAI says textGrain did as well as or better than other methods such as SynthID for text. It warns that good results in testing do not promise solid results in daily use. Detection works about 80 percent of the time for shorter text. Math is harder to detect. Editing the text lowers the success rate.
The watermark will be off by default except for eligible ChatGPT and Codex text in the European Union. Other customers can choose to turn it on for certain models. OpenAI plans to share the technology as open source, so others can study and use the code.
Anthropic did much the same earlier this year for text made by Claude. OpenAI already uses SynthID watermarks and Content Credentials (C2PA) to help people tell if OpenAI made the content. Seven companies including OpenAI and Google promised in 2023 to build watermarking for AI text, images, audio and video. Reuters The European Commission offers icons that companies using AI can add to certain AI-made content. From 2 August 2026, the EU AI Act will require clear labels on AI-made content in key cases. Under the rules, companies using AI cannot depend only on the hidden computer-readable mark.
For teams that build with these models, the weak spots matter as much as how it works. Detection is a matter of odds. Short pieces, math symbols, and normal edits all make it less sure. That matters if a school, publisher, or security team wants to treat a detector result as final proof. Choice-based rollout also means early coverage will be patchy, and limited detector access keeps outside testing small while the system develops.
In my view, this setup fits those limits. A hidden mark from the AI maker, a visible label from the company using the AI, and detector access for researchers spread the work instead of trusting one fix. Sharing textGrain openly could help outside checks and help different detectors work together. The long-term direction is hopeful. Lasting signals that can be tested give platforms, businesses, and regulators something solid to use, even if heavy editing can remove any watermark.


