ChatGPT in the EU will start watermarking texts with an invisible watermark. It won't disappear by copying

Text from ChatGPT can continue to carry a machine-recognizable trace of its origin after being inserted into an email or document. OpenAI announced that it will start adding an invisible watermark to selected text responses from ChatGPT and Codex in the European Union. This is not a visible signature below the response: the mark is encoded directly in the choice of words, and therefore copying it alone will not remove it.

ChatGPT in the EU, developer API voluntarily

Deployment is expected to take place over the coming weeks and will include supported ChatGPT and Codex users across all subscription tiers in the EU. OpenAI is thus responding to the European AI Act’s requirement for generated text to be machine-identifiable. Developers using the API can enable watermarking for selected models globally, but it remains off by default.

Compared to Claude, the implementation is thus geographically narrower. Anthropic opted for global deployment, explaining that it does not yet have a long-term reliable way to restrict watermarking by region. OpenAI, on the other hand, is only enabling automatic watermarking in the EU at launch.

Instead of a signature, a statistical pattern in words

The technology is called textGrain and introduces an invisible statistical signal into word selection. The technical report describes the use of a secret key when ranking candidates for the next word. Individual minor interventions combine in longer text to form a pattern that the detector can recognize only with the help of the text and the corresponding key. The reader will not see any mark.

OpenAI states that in Astra model benchmarks, it did not observe a meaningful difference in performance with and without watermarking. The watermark also does not contain a link to a specific user, account, prompt, or conversation. Detection can thus show a trace of the OpenAI system, but not identify who had the text written.

Short answers and rewriting complicate detection

Success rate heavily depends on length and content. In OpenAI’s tests, at a target false positive rate of 1%, the detector caught the watermark in approximately 80% of passages 200 tokens long and 95% of passages 400 tokens long, for example, from the field of psychology. For mathematics, the results were significantly weaker: a correct answer gives the model less freedom in word choice.

Even more illustrative is a separate text modification test. For passages of 400 tokens, replacing 10% of words with synonyms reduced detection success from approximately 92% to 66%. When 25% of words were replaced, it dropped to 17%. These are OpenAI’s measurement results, not a universal success rate for any document.

Public text verification will not be available yet

The detection result also won’t tell how much work a human put into the text, who owns the text, or whether its claims are true. Even an undetected watermark does not confirm human authorship: obstacles could include short or modified text, an unsupported model, or the use of another AI.

Precisely because of overlooked watermarks and false positives, OpenAI will not make the detector publicly available at launch. Approved researchers and professional organizations can apply for access, and the company will evaluate individual requests separately. Thus, a regular user will not yet receive a public tool into which they can insert text and verify the presence of the textGrain mark.

Would you use a tool to check the origin of text?

Sources: anthropic.com, openai.com, techcrunch.com