Fostering AI alignment through blockchain, proof of personhood and zero knowledge proofs

Abstract

Artificial intelligence (AI) systems are rapidly approaching capabilities that require an increasing level of human control. Existing AI alignment techniques remain opaque, model-specific, and vulnerable in human-level AI, or post-quantum scenarios. To address these issues, this paper proposes a novel AI alignment system architecture in which AI alignment rules are encoded as immutable smart contracts on a blockchain. The blockchain, in turn, is governed by a Proof of Personhood (PoP) consensus mechanism that only admits human agents to the rule validation processes. To protect the privacy of human agents in the identity verification process, the proposed AI alignment system facilitates techniques such as key derivation functions and asymmetric encryption of biometric data. In addition, this system also utilizes blockchain-based decentralized identity (DID) and zero-knowledge proofs (ZKPs). To ensure privacy in post-quantum scenarios, biometric data are linked to zk-STARKs. The proposed AI alignment system is formally described to capture human and AI agents, verification, authentication, and Sybil resistance. The AI shield, a reactive system that prevents unsafe actions by an AI agent that would violate predetermined conditions, enforces the blockchain-based AI alignment rules in real-time, independently of the underlying AI model. Thus, the contribution of this paper is a conceptual framework for the implementation of blockchain technology that utilizes a PoP-based consensus mechanism and zk-STARKs to foster privacy-friendly societal involvement and public auditability of AI developments, providing a democratically governed AI alignment layer applicable to current and future AI models, including those in a post-quantum era.