Just co-authored a whitepaper on a new way to verify AI model execution –without TEEs, ZK proofs, or trusted parties. 

It’s a mechanism for trustless AI inference in decentralized networks 👇

Different models (even subtly different ones) leave statistical fingerprints in their outputs. 

By probing a node with known inputs, you can probabilistically verify what model it’s running –just from an API output :)

This opens the door to decentralized AI networks that are actually verifiable.

- No need for TEEs
- No need for zero-knowledge proofs

Just input/output analysis and crypto-economic incentives to keep nodes honest.

Probabilistic model attestation could be foundational to decentralized AI. 

Whitepaper: https://arxiv.org/abs/2504.13443
DMs open if you’re building in this space. 

–
Special thanks to Michael Yuan for the opportunity to collaborate and James Snewin from for their insightful feedback

The problem:  In decentralized AI networks, how do you know a node is running the right model?

Most proposals rely on trusted hardware or expensive cryptography.  We propose something simpler and more scalable: