OpenAI to roll out invisible watermark for ChatGPT text in the EU
OpenAI will begin invisibly watermarking text generated by ChatGPT and Codex in the European Union to meet the EU AI Act's transparency rules. The method subtly shapes word choices so a detector can identify AI-generated content, though editing can defeat it.
OpenAI announced on Monday that it will roll out an invisible watermark for text generated by ChatGPT and Codex to users across all plans in the European Union over the coming weeks. The move responds to the EU AI Act's transparency requirements, which took effect on August 2 and require companies to mark AI-generated content so other systems can detect it.
According to OpenAI, the watermark is not a visible symbol but works by subtly adjusting the model's word choices, creating a pattern invisible to readers but detectable by a specialized tool. Because the pattern is embedded in the words themselves, it survives when text is copied and pasted elsewhere.
Outside the EU, developers using OpenAI's API anywhere in the world can switch the feature on for select models starting today, though it remains off by default and OpenAI is not making it a global standard at launch.
Method and limitations
Alongside the announcement, OpenAI published a technical report on the method, called textGrain, co-authored with researchers from the University of Pennsylvania and Yale. The approach uses a secret key to sort predictions for the next word in a sequence; combining hundreds of these small nudges lets a detector identify AI-generated text using only the passage and the key.
OpenAI acknowledged the watermark can be removed through editing — in one test, replacing 10% of words with synonyms dropped detection accuracy from about 92% to 66%. Short passages, math answers, and translated text are also harder to detect reliably. Because of these limitations, the company is currently granting detector access only to approved researchers and expert organizations. OpenAI also cautioned that a missing watermark doesn't prove a text was written by a human, since it could be too short, heavily edited, or produced by a rival's AI system.
The announcement comes two months after Anthropic began watermarking text from its Claude model worldwide, a decision that drew backlash from some users. The Wall Street Journal reported in 2024 that OpenAI had built a similar watermarking tool earlier but held off releasing it, partly fearing users would switch to competitors that didn't watermark their output. Anthropic, Google, Meta, Microsoft, and OpenAI are among the companies that have committed to following the EU's code of practice on AI-generated content.
