Some AI writing is about to carry a hidden signature. OpenAI said Monday it will add an invisible watermark, called textGrain, to text from ChatGPT and Codex for eligible users in the European Union over the coming weeks. The goal is to meet the EU's AI Act, which requires AI output to be marked in a way machines can read.

The watermark works by nudging word choices. A language model writes one word at a time from a list of likely options. TextGrain tilts some of those picks, often toward a synonym, leaving a statistical pattern a detector can spot. OpenAI published benchmark scores showing similar results with and without the mark.

The rollout is limited. OpenAI said it is "not making text watermarking a global default at launch." Starting Monday, developers and businesses anywhere that use OpenAI's API can opt in for some models. The detector is not public. Only approved researchers and expert groups can apply, because of the risk of missed marks and false alarms.

The Register, a tech news site, dug into OpenAI's paper and found soft spots. In 200-word passages, the detector may catch only about 80% of watermarks, compared with 95% at 400 words. It does worse on math and code. In tests on 400-token passages, swapping 10% of the words with synonyms cut detection from 92% to 66%. Swapping a quarter of the words dropped it to 17%.

OpenAI is candid about the limits. It says a watermark does not prove a human wrote something, check accuracy, show who owns the text or measure how much a person contributed.

Rivals are moving too. Anthropic announced in August that it would use Google DeepMind's SynthID text method, and it applies the mark worldwide, The Register reported. OpenAI says textGrain "matched or exceeded" SynthID. OpenAI already tags AI images, audio and video with other origin signals.

What we know: EU users of ChatGPT and Codex will get watermarked text in the coming weeks.
What we don't: When, or whether, OpenAI will watermark text for U.S. users by default.