Text Watermarking Method Announced for AI Regulation Compliance
In line with the rules introduced by the European Union AI Act, an optional text watermarking system for global API users and an invisible watermark application for ChatGPT outputs in the EU are being launched.
OpenAI has shared its new text watermarking approach developed under the European Union AI Act, announcing that an optional system is being offered for global API users, while an invisible watermarking era begins for ChatGPT and Codex text outputs in the EU region.
Text Watermarking Approach
Content provenance is being developed to help users understand where texts come from, how they were created, or edited. Tools have previously been introduced for visual and audio content, and a text watermarking strategy is now being implemented.
API and ChatGPT Applications
Starting today, global API customers will be able to optionally select the text watermarking feature for specific models. This feature will be kept disabled by default in the API and will be added to ChatGPT and Codex outputs in the European Union in the coming weeks.
Detector and Open-Source Plan
Applications for access to the text watermark detector have opened, and the process will initially be limited to approved researchers. In addition, it is planned to make the technology available as open-source so that it can be further developed by others.
Technological Limitations and Performance
The technology, named textGrain, adds invisible statistical signals to the model's word choices. Evaluations have shown that performance is on par with other approaches, though it has been noted that shorter texts are more difficult to detect.