OpenAI begins watermarking ChatGPT output in Europe to meet EU AI Act requirements

OpenAI will start embedding an invisible watermark in text produced by ChatGPT and Codex for users in the European Union, a move designed to satisfy the transparency obligations of the EU AI Act that took effect on 2 August. According to a company statement on Monday, the rollout will begin in the coming weeks for all eligible EU users across every subscription tier. API developers worldwide can already enable the mechanism on selected models, where it remains off by default. OpenAI says it has no current plan to make watermarking a global default.

The technique, called textGrain, adds no visual marker. Instead it biases the model’s next-word predictions using a secret key. A technical report released alongside the announcement, co-authored with researchers from the University of Pennsylvania and Mila, demonstrates how hundreds of such statistical “nudges” create a pattern a detector can identify from the text and the key alone. OpenAI says the mark survives copy-paste, does not reveal the user’s identity, and has not measurably degraded model performance when active.

The company acknowledges the watermark is fragile. In internal tests, replacing 10% of words with synonyms cut the detection rate from roughly 92% to roughly 66%. Short texts, mathematical responses, and translated text are also harder to detect. Because of those limitations, detector access will initially be granted only to approved researchers and expert organisations to help evaluate reliability and responsible use. OpenAI explicitly warns that the absence of a watermark “does not prove human authorship” — the text may be too short, heavily edited, or come from another company’s model. The mark can only indicate that an OpenAI system generated or processed part of the passage, not how much human judgment, editing, or creativity was involved.

The announcement follows Anthropic’s global watermarking launch for Claude two months ago, a step that drew criticism from users who argued they supplied the prompts, context, and decisions while the model was merely “the tool.” A 2024 Wall Street Journal report said OpenAI had held similar technology for some time but declined to release it, partly out of concern that users would defect to competitors that did not watermark. Under European regulatory pressure, the company is now deploying the mechanism only in the jurisdiction that requires it. Anthropic, Google, Meta, Microsoft, and OpenAI itself have all committed to the EU Code of Practice on AI-generated content.