Computer scientist Naci Cankaya has published a working draft proposal on LessWrong for verifying that rival nations comply with AI development agreements—think arms-control inspectors for AI compute. The system uses network monitoring and cryptographic checksums to track what runs on AI hardware, while keeping trade secrets and customer data private. According to the proposal, it retrofits to existing hardware without needing a single trusted chip, instead relying on physical security and redundant verification. Cankaya is intentionally publishing this as a draft to invite criticism from security and policy experts before deployment. It's a direct stab at AI governance's hardest problem: verifying what an AI system is actually doing when no one trusts anyone else.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton