The Core Problem: Too Many Decisions, Too Little Attention
Ethereum co-founder Vitalik Buterin has pointed to what he sees as a fundamental issue with decentralized autonomous organizations and democratic governance systems. The problem isn’t about technology or voting mechanisms, but something more basic: human attention.
Writing on social media, Buterin explained that participants in these systems face thousands of decisions across multiple specialized areas. Most people simply don’t have the time or expertise to properly evaluate everything they’re supposed to vote on. I think this is something many of us have felt when trying to participate in governance—there’s just too much to keep up with.
The usual fix has been delegation. You pick someone you trust and let them vote for you. But Buterin argues this creates its own problems. Once you delegate, you’re essentially disempowered. A small group ends up controlling decision-making while everyone else has no real influence after clicking that delegate button.
AI as a Possible Solution
Buterin’s proposed solution involves personal large language models. He shared four different approaches that could help address the attention problem. The first is personal governance agents—these would handle all necessary voting based on preferences learned from your writing, conversations, and direct statements.
When the agent isn’t sure how you’d want to vote on something important, it would ask you directly. But it would provide all the relevant context first, so you don’t have to do all the research yourself. This seems practical, though I wonder about the implementation challenges.
Buterin was careful to distinguish this from what he called dystopian “AI becomes the government” scenarios. When AI is weak, that leads to poor decisions. When AI becomes strong, well, that could be problematic in other ways. But used properly, AI might actually empower people and expand what’s possible with decentralized governance.
Other Approaches: Conversation and Markets
The second approach involves public conversation agents. These would gather information from many participants before giving each person or their AI agent a chance to respond. The system would summarize individual views, convert them into shareable formats without exposing private information, and find common ground between different inputs.
Buterin made an interesting point here: good decisions don’t come from simply averaging people’s views when those views are based only on their own limited information. The process needs to aggregate collective information first, then allow informed responses. That’s a subtle but important distinction.
Suggestion markets represent a third approach. These would create financial incentives for surfacing valuable proposals. Anyone could submit ideas, and AI agents could bet on tokens. When the system accepts an input, it pays out to token holders. This market structure might encourage higher-quality contributions.
Handling Sensitive Decisions
Decentralized governance often struggles with decisions that require secret information—things like adversarial conflicts, internal disputes, or compensation discussions. Organizations typically handle these by giving a few individuals significant power.
Buterin suggested multi-party computation using trusted execution environments as a solution. You’d submit your personal AI into a secure black box. The AI would see private information, make a judgment based on that, and output only the judgment—not the sensitive data itself.
Privacy protection becomes increasingly important as participants submit more personal information. Buterin noted that anonymity needs zero-knowledge proofs, and he believes these should be built into all governance tools from the start.
What strikes me about all this is how much it’s about human limitations rather than technical ones. The technology exists, or will exist soon. The real challenge is designing systems that work with how people actually are—with limited time, attention, and expertise. Buterin’s proposals try to address that gap, though whether they’ll work in practice remains to be seen.
