OpenAI said on Monday in a blog post that, in order to comply with the European Union Artificial Intelligence Act, it will begin adding an invisible watermark to text generated by ChatGPT and Codex within the European Union.
The transparency rules of the EU AI Act, which came into force on August 2 (11 Mordad), require AI companies to mark AI-generated content in a way that other systems can detect it.
OpenAI said this watermark will be activated in the coming weeks for eligible ChatGPT and Codex users on all plans, but only in the European Union. Developers using the OpenAI API anywhere in the world can enable it for selected models starting today; the feature is disabled by default. OpenAI said it will not make the text watermark a global default at launch.
The watermark is not a visible symbol, but by subtly shaping the model’s word choices it creates a pattern that readers cannot see but a detector can recognize. Because the watermark exists in the words themselves, it travels with the text when copied and pasted. OpenAI said the watermark does not identify the user, and that it has not observed any meaningful changes in the performance of its models with it enabled.
Alongside the announcement, OpenAI also published a technical report for its method called textGrain. The report, written in collaboration with researchers at the University of Pennsylvania and Yale, describes an example of using a secret key to order next-word predictions to complete a sentence. Add together hundreds of such signals and a detector can identify AI-generated content using only the text and the key.
Can the watermark be removed by editing? OpenAI’s experiments show that yes. In one test, replacing 10% of words with synonyms reduced detection from about 92% to 66%. The company also said that detecting short texts, math answers, and translated text is more difficult.
The company said: "These limitations factor into our decision to provide initial access to the detector only to researchers and vetted specialist organizations who can help us assess reliability and responsible uses."
OpenAI also warned that the absence of a watermark "does not prove human authorship." Text can be very short or heavily edited, or it might come from another company’s AI.
The company said: "Watermarks can indicate that an OpenAI system generated or processed part of a text, but they cannot indicate how much human judgment, editing, or creativity went into it."
According to TechCrunch, the announcement comes two months after Anthropic said it would watermark text generated by Claude, a move that will be applied worldwide. That decision drew negative reactions from some Claude users who argued they had provided the "prompts, context, and decisions" while Claude was only the "tool."
The Wall Street Journal reported in 2024 that OpenAI had previously built a text watermark but delayed its rollout, partly because it worried users would switch to competitors without watermarks.
Companies such as Anthropic, Google, Meta, Microsoft, and OpenAI are among those that have committed to following the EU’s rules on AI-generated content.