AI-generated text watermarking and detection system
Embed a secret statistical signal into language-model output without materially changing its apparent quality, then use the secret key to detect whether later text was generated through the watermarking process.
From Computerphile — How Watermarks Track AI Generated Content - Computerphile at 02:56
Problem: Determines whether text, code, essays, stories, or other language-model output was produced by a participating AI model, including after some editing.
For: AI companies and organizations that need to identify output generated by particular language models, such as schools or other institutions checking submitted text.
Products from this video
Examples
- Anthropic said it was going to watermark its output.