OpenAI to watermark ChatGPT and Codex text in the EU

OpenAI to watermark ChatGPT and Codex text in the EU

5 October 2026: OpenAI will add an invisible statistical watermark to eligible ChatGPT and Codex text output in the European Union over the coming weeks, and API developers anywhere can opt in to the same watermark from now. The company's own tests show that ordinary editing weakens detection quickly.

Not yet independently verified. We could not open OpenAI's own page, so its statements and test figures come from two outlets that quote it. We will check them against the page when it is reachable. We will update this when it can be confirmed, and remove this note.

The technique is called textGrain. It nudges the model's word choices so that the text carries a pattern a detector can later test for, without changing what a reader sees or copies. OpenAI said: "Over the coming weeks, we will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union." The Register reads the move as OpenAI meeting the letter of the EU AI Act, which requires generative AI providers to make output identifiable by machine.

Outside the EU the watermark is not a default. API customers can switch it on for supported models, and it stays off unless they do.

The detector will not be public. OpenAI is taking applications for access, initially limited to approved researchers and expert organisations.

OpenAI is open about the limits. In tests on 400-token passages, replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%. At a 1% false-positive target, the detector found the watermark in about 80% of 200-token psychology answers, against roughly 95% at 400 tokens. Maths and code do worse because the model has less freedom in its word choice. In OpenAI's words, "The absence of a detected watermark does not prove human authorship."

OpenAI also says a detected mark does not reveal who generated the text, their account, prompt or conversation, and cannot show how much a person later wrote or edited. It says benchmark results for GPT-6 Astra stay broadly similar with textGrain enabled.

Sources

• OpenAI: EU text provenance (page quoted by both outlets below; it returned an access error to our fetch), 5 October 2026
• BleepingComputer: OpenAI is adding invisible watermarks to ChatGPT and Codex text in the EU, 5 October 2026
• The Register: OpenAI rolls out weak sauce watermarking for AI text, 6 October 2026