Buterin sees AI voting tools for DAOs

Buterin said decentralized autonomous organizations often ask participants to evaluate a high volume of proposals that require time and domain expertise, which many holders don’t have. He argued that the typical fallback – delegating votes to a small number of representatives – can concentrate power because supporters “hit the delegate button” and then have little influence over subsequent decisions.

In his proposal, Buterin said a “personal agent” built on large language models could absorb governance context and execute votes according to a user’s inferred preferences, drawn from a person’s writing, conversation history and direct statements. He added that when an agent is uncertain and the issue is important, it should ask the user directly and provide the relevant context needed to make the decision.

The idea targets a participation gap that has become typical across many DAOs. An estimate cited in the same report put average DAO voting participation at roughly 15% to 25% of token holders, a dynamic that can increase the risk of centralization and, in extreme cases, make “governance attacks” easier if a large holder can push through a proposal while others are disengaged.

Buterin also flagged privacy as a constraint for highly decentralized decision-making, particularly when outcomes hinge on sensitive information such as negotiations, internal disputes or funding choices. He suggested that one path could involve participants submitting a personal AI model “into a black box” that can review private information and output only a judgment, without revealing underlying details to other voters. He said larger, more personalized inputs would make privacy protections more important as governance becomes more data-driven.

Work on “digital twin” voting systems has already been discussed inside crypto governance circles. The Near Foundation has previously described plans to develop AI-powered “delegates” that can eventually vote on behalf of DAO members, with a staged rollout moving from chatbot-like assistants to models that represent groups and later individual participants, according to comments from Near Foundation researcher Lane Rettig.

Rettig described an end-state where a user’s “digital twin” learns preferences and acts accordingly at voting time, while also “nudging” users when proposals are relevant to them. He said human oversight would likely remain important for certain categories of decisions, such as major funding allocations or strategic pivots, and argued there should be “a human in the loop” for high-stakes calls.

Near’s governance tooling has also included an AI product called Pulse that tracks community sentiment and summarizes discussion forums, Rettig said, framing it as an example of how early-stage AI can reduce the reading burden for participants before moving toward more autonomous delegates.

Buterin’s remarks come as DAOs continue to search for governance designs that can scale beyond small, highly engaged groups. In practice, many token holders do not read long forum threads or evaluate complex parameter changes, creating a split between nominal decentralization and the smaller set of participants who consistently vote. Buterin’s framing centers on attention as a bottleneck – not just for DAOs, but for democratic decision-making more broadly – and positions personal AI assistants as a way to keep participation broad without forcing every individual to manually process every proposal.