OpenAI starts weaving invisible watermarks into ChatGPT text

OpenAI has begun adding an invisible watermark to text produced by ChatGPT and Codex, and it is starting in the European Union. The company says the mark, called textGrain, will reach eligible users on all plans "over the coming weeks", and that it is not yet a global default.

A mark that lives in the wording

textGrain is not a symbol bolted onto the end of a paragraph. It works by nudging the model's word choices so that the finished text carries a statistical pattern a detector can find and a reader cannot see. Because the pattern sits in the words themselves, it survives being copied and pasted. OpenAI says the watermark does not identify the user, and that it recorded no meaningful change in model performance with it switched on.

The EU AI Act is the trigger

The prompt is regulatory. The European Union's AI Act transparency obligations, which took effect on 2 August 2026, require providers to mark AI-generated content in a way other systems can identify. Alongside the EU rollout, OpenAI has opened watermarking as an opt-in for API customers worldwide on selected models, off by default, and says it is working with cloud partners to offer it through their services in the coming weeks.

What the watermark does not prove

Approved researchers and expert organisations can apply for detector access from today, granted case by case, and the tool reports only whether it finds an OpenAI watermark — not who wrote the text or what they asked. OpenAI is careful about the limits: it says textGrain "does not guarantee reliable detection", does not verify accuracy, does not establish ownership and does not measure how much of a text a person wrote. The method is set out in a technical report, textGrain: entropy-calibrated watermarking for language model text, and OpenAI says it matched or exceeded approaches such as Google DeepMind's SynthID for text.

Photo by Tima Miroshnichenko on Pexels

Our opinion

Watermarking text is a much harder problem than watermarking an image, because language is low-bandwidth: there are only so many word choices a model can nudge before the meaning starts to bend, and every nudge is something a determined user can try to wash out with a paraphrase. OpenAI's own caveats read less like modesty than a warning that this is a provenance hint, not a receipt. The EU-first, opt-in-by-default shape is the tell — a company convinced its mark were robust would not be so careful about where it turns it on.