Why zkML? Because @Microsoft has begun rolling out Copilot Local Actions — letting AI directly perform tasks on your device (system settings, file operations, app control, etc.). Once the AI acts, not just suggests, the question changes: How do we verify that the action matched the user’s intent?

2/ Local execution changes the trust model. There are no server logs. No external oversight layer. The model interprets context → decides → executes. You see the result, but not the reasoning. And when reasoning is invisible, control becomes assumption.

3/ This is where zkML matters: - Prove which model executed the action - Prove the reasoning path followed defined policy - Prove no unauthorized steps were taken All without exposing internal state or system data.

4/ Imagine Copilot where: ✅Every automated action carries a verifiable execution trace ✅ Admins can audit behavior without accessing user content ✅ Users don’t need to “trust” AI — they can verify it behaved as intended zkML turns automation from faith → to assurance.

5/ AI is becoming a system actor — not just an assistant. When intelligence executes directly on the device, proof becomes the safety layer. That’s what @PolyhedraZK is building: verifiable intelligence where action and intent stay aligned.