TLDR:

• Vitalik Buterin argues AI used correctly can empower democratic governance rather than centralize control over it.
  
• Personal AI agents could vote on a user’s behalf by learning from their writing, history, and stated preferences.
  
• Public conversation tools can aggregate views across many participants before asking them to weigh in on decisions.
  
• Multi-party computation allows private governance decisions without exposing sensitive data to any single participant.

AI governance is at the center of a fresh discussion sparked by Ethereum co-founder Vitalik Buterin. He argues that AI, when applied correctly, can push democratic and decentralized governance forward rather than replace it.

His post addresses a long-standing problem: most people lack the time to participate meaningfully in governance decisions.

With thousands of choices across many domains, the current model of delegation concentrates power in too few hands.

Personal AI Agents Could Reshape How People Vote

Buterin proposes using personal large language models to handle the attention problem in decentralized governance.

A personal governance agent could cast votes on a user’s behalf by studying their writing, conversations, and stated preferences. This approach keeps individuals connected to decision-making without requiring constant attention.

When an agent is unsure how a person would vote on a given issue, it would pause and ask them directly. It would also provide all relevant context before prompting any response. This design avoids blind delegation and keeps the individual informed on matters that count.

The model differs sharply from current delegation systems, where supporters often lose influence after pressing a single button.

A personal agent maintains ongoing alignment with the user’s values. It acts as a filter rather than a replacement for human judgment.

Public Conversation Tools Can Aggregate Views More Accurately

Buterin also raises concerns about how collective decisions are currently formed. Simply averaging people’s views based on their own limited information does not produce well-informed outcomes.

A better process would gather and combine information across many participants before asking them to respond.

He points to tools like LLM-enhanced versions of pol.is as one direction worth pursuing. These systems summarize what people have in common based on their actual words. They can surface shared ground that might otherwise stay hidden in large groups.

Additionally, a public conversation agent could translate a person’s views into a shareable format without exposing private details.

This makes broader participation possible without forcing individuals to be publicly identifiable. Anonymity tools using zero-knowledge proofs could support this further.

Multi-Party Computation Addresses Private Decision-Making

One major weakness of democratic governance is its struggle with confidential information. Negotiations, internal disputes, and compensation decisions often require secrecy that open voting cannot provide. Buterin suggests multi-party computation as a technical solution to this tension.

Under this model, a participant’s personal LLM would enter a secure environment, review private data, and output only a judgment.

Neither the participant nor anyone else would see the private information itself. Trusted Execution Environments, or TEEs, have already demonstrated this approach in practice.

Buterin also calls for greater use of garbled circuits to achieve pure cryptographic security in at least two-party cases.

Privacy, he notes, must cover both participant anonymity and the contents of their inputs. Zero-knowledge proofs and multi-party techniques together form the foundation he envisions for this system.