Ethereum co-founder Vitalik Buterin has laid out a governance model that would use personal AI agents to handle voting in DAOs, arguing that large language models could help users cope with the sheer number of decisions decentralized organizations generate. In his proposal, an agent would vote based on a person’s writing, conversations, and stated preferences. If the system cannot confidently infer the user’s position on an important matter, it would ask the user directly.
His argument is aimed at a familiar problem in onchain governance: too many decisions, too little attention, and a tendency for influence to drift toward a small set of delegates. Buterin said personal AI agents could reduce that concentration by helping participants stay involved without forcing them to review every proposal by hand. The point is not to remove people from the process. It is to make their judgment usable at scale.
Personal agents as a layer above standard delegation
Under this approach, a personal AI agent would stay aligned with a user’s values over time and filter governance matters according to relevance and importance. That changes the mechanics of participation. Instead of giving away influence through a standard delegation model and stepping back, users could keep a more active role across many decisions, while still being pulled in when an issue carries weight or the agent is uncertain.
Buterin framed this as a way to address both attention limits and expertise limits inside decentralized organizations. Many DAO structures are open in theory but difficult to participate in consistently. An agent that tracks preferences and handles routine choices could lower the cognitive burden and make broad participation more practical.
Public conversation agents and zk-based privacy
Buterin also discussed how information could be aggregated across groups without exposing private inputs. He proposed public conversation agents that summarize shared themes from participants’ contributions while keeping underlying data private. With LLM support, those systems could convert personal views into formats that are easier to share and deliberate on, without stripping away anonymity.
Zero-knowledge proofs would be part of that design, helping secure participant identities during discussion and coordination. The model is meant to improve on simple linear voting, which can miss the distributed knowledge spread across a community. Instead of reducing governance to raw yes-or-no tallies, AI agents could respond to synthesized group insights and help build decisions from a richer information base.
That creates a bridge between private opinion and collective deliberation. Individual participants do not have to reveal everything they know, but their input can still shape a broader summary that others can use.
MPC and TEEs for sensitive governance decisions
For cases involving confidential information, Buterin pointed to multi-party computation and trusted execution environments, or TEEs, as possible security layers. In that setup, personal AI agents could process sensitive inputs inside protected environments and release only the final decision. The underlying data would remain hidden.
He said this structure could be applied to negotiations, disputes, and compensation decisions, where both identity privacy and content privacy matter. The proposal combines three components: personal AI agents for individual preference expression, public conversation agents for group-level synthesis, and cryptographic systems for privacy protection. Together, Buterin presented them as a possible framework for scaling democratic governance in decentralized systems.