OpenAI will watermark ChatGPT and Codex text in the EU

OpenAI said on 5 October that it is expanding its approach to content provenance to include text, "in response to EU regulatory requirements", while acknowledging "significant limitations of current text watermarking technology". In a thread on X, the company said its tools already help verify whether an image or audio file was created with its products. ChatGPT and Codex output in the EU is next, over the coming weeks.

OpenAI describes the mechanism as embedding an invisible statistical signal in text as it is generated. It says a watermark helps answer one question: whether a passage was likely generated by an OpenAI model. TestingCatalog reports that the watermark will stay off by default in the API. The Decoder reports that API customers worldwide can opt in, which it contrasts with Anthropic.

OpenAI is open about how fragile the signal is. "Watermarks have limits. They're often undetectable, especially in short passages," OpenAI wrote, adding that rewriting or translating text can remove the watermark entirely. For now, only approved researchers will get access to its detector, so they can help evaluate and improve the technology, according to the company.

The Decoder, citing OpenAI's own tests, reports detection rates as high as 95 percent that fall to 17 percent when a quarter of the words are replaced. We could not open OpenAI's full post, so those figures come from The Decoder's account and have not been checked against the original. The Decoder calls the watermark textGrain.

Reaction on X was mixed. Andrew Curran noted that, unlike Anthropic and Google, OpenAI is launching its watermark only in the EU rather than globally. The newsletter writer kimmonismus said he does not like the direction, because the pattern is embedded in the wording ChatGPT chooses. Neither claim has been confirmed by OpenAI beyond its own thread.