Language Model Output
Core concept regarding the generation, structure, and provenance of text produced by large-language-models.
Content Provenance & Watermarking
Recent developments in AI ethics focus on embedding verifiable signals in Language Model Output to distinguish synthetic from human-generated text.
- Invisible Fingerprints: Claude AI has implemented an invisible text watermarking system that leaves subtle, undetectable patterns in its generated content Claude AI Invisible Text Watermarking Report.
- Distinction from Visual Watermarks: Unlike traditional image watermarks, this method operates at the token level, affecting the statistical distribution of output without altering readability.
- Broadening Scope: The application of watermarking extends beyond text to distinguishing AI-generated content from natural biological sequences and deepfakes, addressing challenges across multiple modalities AI Watermarking: Distinguishing AI-Generated Content and Biological Sequences.
- Key Research: Google DeepMind explores the science of watermarking for both deepfakes and DNA, highlighting the growing need for provenance verification as generative AI capabilities advance.