OpenAI will begin adding an invisible watermark to text created by ChatGPT and Codex in the European Union. The company announced the plan on Monday in a blog post.
The move is meant to help OpenAI comply with the EU AI Act. The law's transparency rules took effect on August 2.
Those rules require AI companies to mark AI-generated content in a way that other systems can identify.
How the Watermark Works
The watermark is not a visible symbol. Instead, it subtly shapes the words the model chooses, leaving a pattern that readers cannot see but a detector can find.
Because the pattern lives in the words themselves, it stays with the text when it is copied and pasted. OpenAI said the watermark does not identify the user.
The company also said it saw no meaningful change in how its models performed with the watermark switched on.
OpenAI released a technical report on the method, which it calls textGrain. The report was co-written with researchers from the University of Pennsylvania and Yale.
It explains how a secret key is used to sort next-word predictions as the model finishes a sentence. Hundreds of these small nudges add up, letting a detector spot AI-generated text using only the text and the key.
Limits of the System
OpenAI's own tests show the watermark can be weakened by editing. In one test, swapping 10% of words for synonyms lowered detection from about 92% to 66%.
The company said short passages, math answers, and translated text are also harder to detect.
Because of these limits, OpenAI will at first give detector access only to approved researchers and expert organizations. The company said these groups can help it judge reliability and responsible uses.
OpenAI also warned that a missing watermark "does not prove human authorship." The text could be too short, heavily edited, or written by another company's AI.
The company said a watermark can show that an OpenAI system produced or processed part of a passage. It cannot show how much human judgment, editing, or creativity went into it.
The announcement comes two months after Anthropic said it would watermark text from its Claude models worldwide. Some Claude users pushed back, saying they supplied "the instructions, context, decisions" while Claude was just "the tool."
OpenAI had built a text watermark before but did not release it. The Wall Street Journal reported in 2024 that the company worried users might switch to rivals that did not watermark.
Anthropic, Google, Meta, Microsoft, and OpenAI are among the companies that have committed to the EU's code of practice on AI-generated content.
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 it is not making text watermarking a global default at launch.
Developers using OpenAI's API anywhere in the world can turn the watermark on for select models as of Monday. The feature is off by default.
