The Network State Podcast #37 - Uma Roy
60 snips
Apr 23, 2026 
Uma Roy, cryptographer and cofounder of Succinct, builds zero-knowledge virtual machines to prove digital media provenance. She explains zero-knowledge proofs, how cryptography can defend against AI-driven fakes, and why on-chain proofs and signed edit chains can restore trust in images and text. The conversation covers detector failures, device-based signing, and practical stacks for provable media.
AI Snips
Chapters
Transcript
Episode notes
Zero Knowledge As A Defense Against AI Fakes
- Zero knowledge proofs can act as a defense against AI-generated fakes by proving attributes without revealing private data.
- Uma Roy describes Succinct's ZKVM as a foundation-model-like layer that makes complex ZK statements easy to produce and verify.
AI Detection Is Fragile To Simple Perturbations
- AI detectors trained to spot model outputs are fragile and break under tiny perturbations like blur or Gaussian noise.
- Uma Roy's Succinct benchmark showed commercial image detectors drop from tens of percent to single digits after trivial image perturbations.
Adopt No Public Undisclosed AI Policy
- Adopt a policy of no public undisclosed AI to maintain trust in public-facing content.
- Balaji explains public disclosed AI is acceptable, private AI is fine, but undisclosed AI causes audience distrust and tax on the verifier.