Technology

ChatGPT in Europe Is Getting an Invisible Watermark

Martin HollowayPublished 2month ago3 min readBased on 4 sources
Reading level
ChatGPT in Europe Is Getting an Invisible Watermark
source:openai.com

OpenAI is adding a hidden watermark to text written by ChatGPT and Codex for users in the European Union.

The company announced the plan on October 5, 2026, in a post titled "Our approach to EU text provenance rules" OpenAI. The system is called textGrain. As the model writes, it chooses between likely next words. textGrain slightly favors some choices over others, according to OpenAI's help article titled "Provenance signals in OpenAI-generated content" OpenAI. Think of it like a faint pattern pressed into paper that you cannot see, but a special tool can find. The text still reads normally, with no label or warning added.

All ChatGPT and Codex plans in the EU will get watermarked text over the coming weeks. OpenAI is not turning this on for the rest of the world at launch The Verge.

Customers who use OpenAI's API have a different option. They can choose to turn on watermarks for certain models, anywhere in the world. That choice matters because even small changes in word choices can affect testing and tools that process the text.

Only certain people can check for the watermark. Approved researchers and expert groups can apply to use OpenAI's detector, with early access decided case by case. The detector only says if it sees an OpenAI watermark. It does not name the user or show prompts or chats.

OpenAI says there are clear limits. textGrain does not always work reliably. A watermark does not show if the text is correct, who owns it, how much a person helped write it, or if a person wrote it. Watermarked and normal text scored about the same in OpenAI's tests.

The watermark is built into the word choices themselves, not added as extra data or formatting. So it survives copy and paste. But it can break down if the text is reworded, translated, heavily edited, or written partly by a person and partly by AI. It also works less well on short answers, on very strict text like code, and on text rewritten by another AI model. Longer, more open writing keeps a stronger signal.

The broader context here is regional rules. OpenAI described this as following EU AI Act rules. Starting in one region, with an optional global switch for developers and limited detector access, gives regulators something computers can check without changing everything at once. It also keeps the detector private for now, which makes it harder to cheat while researchers test how often it is wrong.

In my view, it helps to think of textGrain as background plumbing, not as a lie detector. Because it is hidden, based on odds, and not proof of who wrote something, it will not settle arguments about authorship. It can help platforms, auditors, and researchers sort large amounts of text and improve testing tools. The danger is if bosses, teachers, or courts treat it as a simple yes-or-no answer. OpenAI's warnings about reliability, correctness, ownership, and human help should be repeated wherever a detector result is used.

In practical terms, what this enables is a careful test. EU use will show effects on quality, speed, complaints, and workarounds. The API option allows comparison between watermarked and normal text on the same models. Limited detector access allows outside checks before wider release. Teams building with these models should record if watermarking was on, watch for changes in output, and treat any detector result as one weak clue, not the final answer.