Safety & Alignment
Watermarking
Embedding statistical patterns into model outputs so they can be detected as AI-generated.
Definition
Watermarking biases token sampling toward a secret 'green list' of tokens, leaving a statistical signature detectable to anyone with the key. Image watermarks (SynthID) embed signals in pixel space. Adoption is uneven and active research explores robustness to paraphrasing.
Common use cases
- Provenance
- AI detection
- Content authentication