OpenAI Adds Invisible Watermarks to AI Text in the EU

OpenAI has begun adding invisible, machine-readable watermarks to text produced by some of its models, starting with users in the European Union. The company said the move is meant to align with the EU AI Act’s transparency requirements, which ask providers to make AI-generated content identifiable. The feature is being switched on in the EU over the coming weeks, and from now on global API customers can also choose to turn it on for specific models.

The watermarking technique is called textGrain. Rather than stamping a visible label on a paragraph, it works during the model’s word-selection process, weaving a statistical signal into the choices the model makes as it writes. A detector that knows what to look for can read that signal and decide whether a stretch of text carries an OpenAI watermark. To a normal reader, the text looks unchanged.

That subtlety is the point. Visible labels are easy to strip: a user can copy a paragraph, ask another model to rewrite it, or simply delete a footer note. A hidden statistical trace is harder to remove without altering the text itself. It also avoids the awkwardness of interrupting a piece of writing with a banner that says a machine wrote it, which can be intrusive in a chat, a draft email or a line of code.

Watermarking is nonetheless a technical compromise rather than a perfect seal. Detection usually needs access to the right tool, and signals can weaken when text is heavily edited, translated or mixed with human writing. OpenAI has framed textGrain as one layer in a broader transparency effort rather than a guarantee that every AI-assisted sentence can be traced. The company has not said that the marking will apply to every model or every region, and the EU rollout is described as applying to qualifying ChatGPT and Codex output.

The EU dimension matters because the bloc’s AI rules are among the first to require this kind of content transparency from major model providers. Companies that serve European users have been adjusting products, documentation and internal processes to match. For OpenAI, adding watermarks is both a compliance step and a way to test how the technology behaves at the scale of a mainstream chatbot used by hundreds of millions of people.

For API customers, the choice to enable watermarking on selected models gives businesses a lever. A publisher might want AI-assisted text marked so its own systems can flag it later. A developer building a writing tool might prefer to leave it off, depending on the audience. Because the setting is optional for API users outside the EU, adoption will likely vary by industry and by how much a company cares about provenance.

Watermarking has become one of the more active corners of AI research, with labs, universities and standards bodies debating how to mark synthetic media without breaking it. Text is harder than images: a picture can carry hidden pixel patterns, but words are discrete and easily rearranged. Approaches like textGrain try to make the mark survive normal use while staying invisible to readers.

Whether invisible watermarks become a default expectation or fade into a niche compliance tool will depend on how well detection holds up in the wild. For now, OpenAI’s EU rollout gives the idea its largest real-world test yet, and gives global developers an early switch to experiment with.