OpenAI has confirmed it will start automatically adding an invisible watermark to text produced by ChatGPT and Codex, but only for users in the European Union. The feature, built around a proprietary technique called textGrain, is the company’s answer to new EU transparency rules on AI-generated content.
Key takeaways
OpenAI’s watermark will be on by default only for ChatGPT and Codex users inside the EU, rolling out over the coming weeks.
Outside the EU, including through the API, the feature exists but stays switched off unless a developer turns it on.
The textGrain method reaches about 92% detection accuracy, but editing just 10% of the words can cut that to around 66%.
OpenAI is limiting watermark detector access to approved researchers and expert organizations for now.
Rival Anthropic rolled out watermarking globally in August, unlike OpenAI’s region-specific approach.
Why OpenAI is watermarking ChatGPT text only in the EU
OpenAI’s default watermarking applies exclusively to European users for now, with no global switch planned at launch. The company said in a blog post, reported by TechCrunch, that it is “not making text watermarking a global default at launch,” even as it opens the feature to developers worldwide through its API starting this week, off by default. As Ars Technica reported, the rollout to eligible ChatGPT and Codex accounts on all plans in the EU will happen “in the coming weeks.”
BleepingComputer cited OpenAI’s own wording from the announcement: “Over the coming weeks, we will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union.” The watermark itself is invisible to anyone reading or copying the text, and OpenAI says it travels with the content since it lives in the word choices themselves rather than in a visible tag.
The EU AI Act behind the compliance push
The driving force is the EU AI Act’s transparency obligations, which took effect on August 2, 2026, according to TechCrunch. The rules require AI companies to mark machine-generated content in a way that other systems can detect. Standards such as SynthID and the C2PA project already exist for this purpose, but Ars Technica noted they remain relatively easy for anyone with basic technical know-how to circumvent.
How textGrain works, and where it falls short
OpenAI’s method nudges the model’s word choices in subtle, statistically detectable patterns that a specialized detector can read using a secret key, without changing the output’s apparent quality to a human reader. The company published a technical paper on textGrain co-written with researchers from the University of Pennsylvania and Yale, describing how the key sorts next-word predictions to shape each sentence.
The limits of that approach are notable. Tests cited by Ars Technica and TechCrunch show roughly 92% detection accuracy on unedited text, but changing about 10% of the words can cut detection to around 66%, and altering 20% of the text can reduce the success rate by nearly 75%. Short passages, math answers, and translated text are also harder to detect accurately, OpenAI acknowledged.
Because of these gaps, OpenAI said it is “provid[ing] initial detector access only to approved researchers and expert organizations, who can help us evaluate reliability and responsible uses,” per the quote carried by TechCrunch. Over time, others may seek approval as well, and OpenAI further warned that the absence of a watermark “does not prove human authorship,” noting that text might be too brief, too heavily revised, or produced by another company’s AI entirely. As the company put it, watermarks “can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it.”
OpenAI’s approach versus Anthropic’s global rollout
The contrast with Anthropic is direct. In August, Anthropic rolled out watermarking for text generated by Claude and activated it worldwide—a move that, according to TechCrunch, sparked pushback from some Claude users who contended they provided “the instructions, context, decisions” while Claude merely served as “the tool.” By contrast, OpenAI currently applies its default setting only where regulations require it, at least for now within ChatGPT and Codex.
According to TechCrunch, OpenAI had previously developed a text watermark but chose not to launch it, partly out of concern that users might migrate to competing tools lacking such a feature, a detail drawn from a 2024 Wall Street Journal report referenced in the piece. Anthropic, Google, Meta, Microsoft, and OpenAI have all committed to the EU’s code of practice on AI-generated content, per TechCrunch.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.