OpenAI will begin automatically watermarking text generated with ChatGPT in the European Union, the company announced. It will also offer the watermark feature in other regions, but it will be disabled by default outside the EU.
The move in Europe is driven by the need to comply with the EU AI Law, which came into force in August. It requires marking content produced by AI models so that another tool can detect it. Unfortunately, there is still no completely effective and reliable way to do this. Some standards already exist, such as SynthID and the C2PA project, but they are relatively easy to circumvent for anyone with basic knowledge.
The same is probably true of the OpenAI watermark. His method is proprietary; the company calls it textGrain and has published a white paper explaining how it works. But generally it works like other LLM watermarking tools we’ve seen in the past: it places patterns in word choices that aren’t clear to a human reader and don’t significantly change the overall quality of the result, but that someone with a key can use a specialized detector to find. OpenAI says it will give access to the detector to a limited number of researchers and organizations, and provide an approval application process for others to be added over time.
