OpenAI announced textGrain on October 5 (US time), a system that quietly leaves a marker in text written by AI. In the EU, it will be applied to ChatGPT and Codex output across all plans over the coming weeks, while API customers anywhere in the world can switch it on voluntarily. The move is driven by the transparency obligations of the EU AI Act, and the key question is how useful a watermark that cannot be seen really is.

The mark lives in word choice, not in characters

textGrain does not insert special characters or spaces into text. When the model picks its next word and several candidates fit about equally well, a secret key nudges the choice slightly in one direction. No single word looks odd, but when a passage of reasonable length is analyzed statistically, the bias shows up.

The benefit is that people cannot see it and it is hard to strip out by copy and paste. The trade-off is that detection requires a dedicated tool.

Who gets it, and when

Coverage depends on how the service is used.

EU users of ChatGPT and Codex will have it applied automatically on all plans over the next several weeks. It is not a global default; it is specific to the EU.

API customers can opt in starting October 5. It is off by default and can be enabled per project or per organization.

The detector is not publicly available. Approved researchers and expert organizations can use it on a case-by-case basis, and applications also opened on October 5. Because watermarks are vulnerable to manipulation, access is kept narrow.

The EU AI Act, Article 50

The trigger is Article 50 of the EU AI Act, which requires providers of generative AI to make output identifiable in a machine-readable way. The transparency rules took effect on August 2 and apply to anyone serving users in the EU.

According to OpenAI, the EU Code of Practice on transparency does not require watermarks for outputs shorter than 200 tokens or for code snippets. Code-focused use is therefore largely unaffected, and the main target is generated prose.

Detection is fragile against editing

OpenAI's figures, at a 1 percent false-positive rate, show detection of about 80 percent for a 200-token passage and about 95 percent for 400 tokens. Those numbers apply to untouched text under ideal conditions.

Accuracy falls sharply once text is altered. Replacing 10 percent of words drops detection from 92 percent to 66 percent, and replacing 25 percent drops it to 17 percent. Translation and heavy editing also make detection harder, as do very short passages and text with little room for variation, such as mathematics. Across EU languages, Spanish reportedly scored highest at 69 percent and Romanian lowest at 42 percent.

These numbers come from OpenAI's own evaluation, and no independent verification has been presented yet.

No watermark does not mean human-written

OpenAI is explicit about the limits. A watermark does not measure how much a human contributed, does not establish ownership or responsibility, does not identify the user, and does not verify accuracy. The absence of a detected watermark does not prove the text was written by a person.

Text from other companies' models, or AI drafts that a person heavily rewrites, tends to fall outside detection. Judging someone based only on a detection result, in schools or publishing, would be risky. Concerns have also been raised that non-native speakers who polish their writing with AI could be wrongly flagged.

Summary

textGrain is an invisible text watermark that began as a response to the EU AI Act. It will be applied step by step to ChatGPT and Codex in the EU, and API customers can opt in from October 5. Because it is weak against rewriting and the detector is available only to a limited group, it is not a universal way to spot AI text. For now, the thing to watch is how this first step in regulatory compliance develops in practice.