OpenAI announced the introduction of text watermarking in a blog post on Monday, Oct. 5. A watermark is a mechanism that places a mark invisible to the human eye in AI-written text so that it can later be determined whether the text came from AI. It covers text generated by ChatGPT and Codex, and the mark is applied only on the screens of EU users. Outside Europe, it is switched off.
The company's stated purpose is compliance with the EU AI Act. The AI Act is a law in which the European Union sets out the obligations AI service providers must meet, and its transparency provisions took effect on Aug. 2. The response came just over two months later.
OpenAI named the technology textGrain. The company claimed that textGrain performs on par with or better than other text watermarks such as Google DeepMind's SynthID, but it did not release the metrics or results used in the comparison. It also said it has begun a limited release of the detection technology that reads the marks. The performance claims are the company's own assessment.
The company disclosed the weakness itself. OpenAI explained that if a person edits the generated text, the watermark may not be reliably detected. In assignments and reports, where reworking paragraphs is common, this weakness becomes the limit of its usefulness.
The company also set the scope of application itself. OpenAI said it is not using textGrain as a global default, and gave room to learn from real-world use and feedback as the reason for turning it on region by region. With the same prompt, an answer received in Berlin carries the mark, while one received in Seoul does not. It is hard to see this as a tool switched on because the technology had fully matured.
The scope of application was drawn to follow where the law applies.
Since OpenAI said it would apply it only in the EU "initially," there is room to expand it to other regions. The company has said nothing about when, or under what conditions, it would expand. Anthropic is also embedding invisible watermarks to comply with EU rules.
A report handed to a university teaching assistant in Brussels may carry a hidden, invisible mark. A report received by a teaching assistant in Busan does not. Even with text written using the same ChatGPT, a trace may be left or may disappear depending on where it was written, so anyone who writes with AI would do well to know which region's rules their writing follows.
