As AI content floods the internet (AI and Misinformation, Deepfakes and Synthetic Media), two approaches try to keep truth legible: watermarking (mark AI output) and provenance (prove where real content came from).
Watermarking — mark the fakes
Embed an invisible, statistical signal in AI output so a detector can later say "this was AI-generated."
| Type | How |
|---|---|
| Text | Bias token choices into a detectable pattern |
| Image/audio | Imperceptible signal in pixels/samples (e.g. SynthID) |
The weakness: watermarks can be weakened by editing/paraphrasing, and open models can omit them entirely. Detection is an arms race, not a solution.
Provenance — prove the reals
Flip it: cryptographically sign authentic content at the source.
- C2PA attaches signed, tamper-evident metadata (who, when, what tools) using Digital Signatures.
- Cameras and editors stamp a verifiable history; altering it breaks the signature.
Rather than chasing every fake, provenance lets real media prove itself — the more durable bet against a sea of synthetic content.
Related: Deepfakes and Synthetic Media · AI and Misinformation · Digital Signatures