OpenAI to watermark ChatGPT text by default in the EU
OpenAI will automatically add a watermark to ChatGPT-generated text in the European Union to meet EU AI Act requirements. Elsewhere the feature will be available but switched off by default.
OpenAI has announced that it will start automatically watermarking text produced by ChatGPT in the European Union. Outside the EU, users will be able to use the feature, but it will have to be turned on manually.
Driven by EU law
The decision stems from the need to comply with the EU AI Act, which took effect in August. The law requires content generated by AI models to be marked in a way that another tool can detect.
However, there is still no fully effective and reliable method for doing so. Several standards already exist, such as SynthID and the C2PA project, but people with basic know-how can circumvent them fairly easily. The same is likely to hold for OpenAI's approach.
How textGrain works
OpenAI's method is proprietary, and the company calls it textGrain. A technical paper describing how it works has been published. In general terms it resembles other language-model watermarking tools seen before: it introduces patterns into word choices that a human reader does not notice and that do not meaningfully reduce the quality of the output. Someone holding a key can nevertheless find them with a specialized detector.
Access to the detector
OpenAI says it will give access to the detector to a limited number of researchers and organizations. Others will be able to be added over time through a request-for-approval process.
