OpenAI will start watermarking ChatGPT texts by default in the European Union
The invisible textGrain watermark will be added to texts from ChatGPT and Codex in the EU within a few weeks. OpenAI is doing this to meet the transparency requirement under the AI Act, but it admits the mark is easy to erase.
OpenAI announced on Monday, October 5, that it will start watermarking text generated by ChatGPT and the developer tool Codex by default for users in the European Union. According to MamStartup, citing TechCrunch, the change will apply to all subscription plans. The watermarking is expected to appear “in the coming weeks,” Ars Technica (opens in a new tab) reports. Outside the EU and in the API (the interface through which companies connect models to their own applications), watermarking is available as an option, turned off by default.
The reason is the AI Act, the EU regulation on artificial intelligence. Article 50, whose transparency provisions have applied since August 2, 2026, requires providers of AI systems to label synthetic content in a machine-readable way. Anthropic introduced similar watermarking for Claude models in August, but enabled it worldwide.
OpenAI’s system is called textGrain. When choosing each subsequent word, the model, based on a secret key and the preceding text, slightly favors some of the possible words. To the reader, the text sounds the same, while a detector with the same key recognizes the statistical pattern.
In the best-case scenario, OpenAI’s tests show a 92 percent detection rate. According to Ars Technica, changing 10 percent of the text lowers it by nearly 30 percent, and changing 20 percent by nearly 75 percent. Spider's Web (opens in a new tab), discussing OpenAI’s charts, writes about replacing 10 and 25 percent of the words. Effectiveness also drops with short and translated texts, and rewriting by another model erases the mark. For now, only a limited group of approved researchers and organizations will get the detector.
[Watermarks] may indicate that an OpenAI system generated or processed part of a text passage, but not how much human judgment, editing, or creativity went into it.
For companies in the EU, this means that texts prepared in ChatGPT by employees and agencies will carry a hidden marker. However, its absence does not prove human authorship, and its presence does not say who edited the text or to what extent.