OpenAI will start adding an invisible watermark to text generated by ChatGPT and Codex for users in the European Union to comply with the EU AI Act. This regulation enforces transparency rules that took effect on August 2, requiring AI-generated content to be identifiable by other systems.
Over the coming weeks, the watermark will be rolled out to eligible ChatGPT and Codex users on all plans within the EU. Additionally, developers using OpenAI’s API globally can enable it for specific models starting today, although it is off by default. OpenAI has clarified that text watermarking will not be a global default at launch.
This watermark doesn’t take the form of a traditional symbol; instead, it subtly influences the model’s word choices, creating a pattern that remains invisible to readers but detectable by appropriate systems. Since it is embedded in the text itself, it travels with the text when copied and pasted. Notably, the watermark does not identify individual users, and initial tests have shown no significant impact on model performance when the watermark is active.
OpenAI also published a technical report for its watermarking method, called textGrain, co-written with researchers from the University of Pennsylvania and Yale. It walks through how a secret key sorts next-word predictions, so a detector can spot AI-generated content using only the text and the key.
There are some limitations, however. OpenAI’s assessments suggest that the watermark can potentially be removed through editing. For instance, one test showed that replacing 10% of words with synonyms dropped the detection rate from about 92% to 66%. Furthermore, shorter passages, mathematical answers, and translations are reportedly more challenging to detect.
OpenAI emphasizes that the absence of a watermark does not necessarily mean a piece of text was written by a human. Factors such as the text’s brevity, excessive editing, or the use of another company’s AI could explain this lack of identification. The company clarified that while watermarks may indicate that content was generated or processed by an OpenAI system, they do not measure the extent of human judgment, editing, or creativity involved.
The announcement comes two months after Anthropic said it would watermark text generated by Claude, a move it is applying worldwide. That decision drew backlash from some Claude users, who argued they had supplied the instructions, context, and decisions while Claude was just the tool.
Interestingly, OpenAI had previously developed a text watermark but decided against releasing it. Concerns about potential user migration to competitors that do not use watermarking contributed to this delay. OpenAI now joins other major players in the industry, including Anthropic, Google, Meta, and Microsoft, all of whom have committed to following the EU’s code of practice regarding AI-generated content.
