OpenAI adds invisible watermarks to ChatGPT text for EU users

OpenAI said Monday it will start adding an invisible watermark, a hidden signal embedded in generated text, to output from ChatGPT and Codex in the European Union. The company said the move is meant to comply with the EU AI Act, whose transparency rules took effect on August 2 and require AI companies to mark AI-generated content so other systems can detect it.

The watermark will roll out over the coming weeks to eligible ChatGPT and Codex users on all plans, but only in the EU, OpenAI said. Developers using OpenAI’s API anywhere in the world can turn the feature on for select models starting today, though it is off by default. OpenAI said it is not making text watermarking a global default at launch.

The watermark has no visible symbol. Instead, it works by subtly shaping which words the model chooses, leaving a pattern invisible to readers but readable by a detector. Because the pattern lives in the words themselves, it survives copying and pasting. OpenAI said the watermark does not identify the individual user, and that turning it on produced no meaningful change in model performance.

Alongside the announcement, OpenAI published a technical report on the method, called textGrain, co-written with researchers from the University of Pennsylvania and Yale. The report describes using a secret key to sort next-word predictions as the model finishes a sentence. Stacking hundreds of these small nudges together lets a detector identify AI-generated text using only the text and the key.

OpenAI’s own testing shows the watermark can be erased through editing. Replacing 10% of words with synonyms dropped detection from about 92% to 66%, the company said. Short passages, math answers, and translated text are also harder to detect. “These limitations contribute to our decision to provide initial detector access only to approved researchers and expert organizations, who can help us evaluate reliability and responsible uses,” OpenAI said. The company also cautioned that a missing watermark does not prove a human wrote the text, since a passage could be too short, too heavily edited, or produced by a different company’s AI. “[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 said.

The announcement follows Anthropic’s move two months ago to watermark text generated by Claude worldwide, a decision that drew pushback from some users who argued they supplied the instructions and judgment while Claude was just a tool. OpenAI had reportedly built a text watermark before but held off releasing it, partly out of concern that users would switch to competitors without one, The Wall Street Journal reported in 2024. Anthropic, Google, Meta, Microsoft, and OpenAI have all committed to the EU’s code of practice on AI-generated content.

What happens next depends on how far OpenAI extends the feature once the EU rollout finishes, and whether detector access expands beyond the approved researchers and expert organizations it is currently limited to.