Vitalik Buterin Proposes AI ‘Stewards’ to Overhaul DAO Governance and Fix Token-Weighted Voting

Markets barely reacted to the idea, which is probably healthy. At the time of the discussion circulating, Ethereum$1,686.33 traded around $1,991, up roughly 1.4%, while Bitcoin$62,588.20 hovered near $68,556, up about 1.35%. Governance reform pitches rarely move spot prices, and that is appropriate: most DAOs do not fail because they lack a clever voting mechanism, they fail because participation is low, incentives are messy, and power concentrates fast. ^[2]

What Buterin is actually proposing

Buterin's latest governance sketch argues that many DAOs need an extra layer between raw token balances and final decisions: AI "stewards" that help interpret proposals, surface tradeoffs, and make recommendations based on an explicitly stated set of goals. ^[3]

Key point: this is not "let the bot run the treasury." The concept is closer to AI as a governance civil servant, where models are trained, constrained, and evaluated against a community's values. Think of it as structured advice plus accountability, not autopilot.

Under the framing Buterin has been pushing, DAOs should aim for systems that are:

• Legible: participants can understand why an outcome happened.
• Resistant to capture: harder for whales, insiders, or bribe markets to dominate.
• Functional under low participation: because most token holders do not vote, and pretending otherwise has not helped.

AI stewards would, in theory, reduce the cognitive cost of participating by summarizing long proposals, comparing them to prior decisions, flagging hidden risks, and scoring them against the DAO's stated priorities.

The problem being targeted: token voting is "working" exactly as designed

Token-weighted voting is simple: one token, one vote (sometimes delegated). The obvious downside is plutocracy. The less obvious downsides are arguably worse in practice:

• Voter apathy: if a few large holders decide outcomes anyway, smaller holders rationally disengage.
• Information asymmetry: governance discussions get technical fast, and most participants do not have time to audit code, read budgets, or model second order effects.
• Bribery and vote markets: where votes become a financial instrument, not an expression of preference.
• Governance theater: forums stay active, but real power sits with a small cluster of delegates, insiders, or treasury gatekeepers.

Buterin's critique, echoed across the recent coverage and commentary, is that DAOs need to be "different and better" rather than simply "more." If governance reliably concentrates power, the system is not failing, it is delivering its default outcome. ^[4]

Where AI stewards fit, and what "steward" even means

The "steward" label matters. A steward is not a sovereign. The value proposition is that AI can act as an interpretation and coordination layer that makes governance more usable without claiming final authority.

A practical implementation could look like this:

1) Proposal triage and compression

Most DAO proposals are long, repetitive, and filled with context links. An AI steward could generate:

• A short summary
• A list of assumptions and missing details
• A risk checklist (security, legal exposure, vendor dependency)
• Budget sanity checks (runway impact, comparable spend, conflicts)

This does not fix plutocracy. It does reduce the "I cannot even parse this" barrier that keeps participation low.

2) Value alignment against a written constitution

The harder part is aligning recommendations with community intent. Buterin has repeatedly emphasized governance designs that rely on explicit objectives, sometimes described as a "constitution," so decision-making is not just vibes plus token balances.

An AI steward could be constrained to reason from that constitution, then cite which clauses or priorities a proposal supports or violates. That moves governance from "who yelled loudest on Discord" to "which principle are we optimizing for."

3) Multi model, adversarial review

One model giving advice is easy to capture or bias. Multiple stewards with different training data, incentive structures, or even competing "parties" could produce adversarial critiques of the same proposal. Humans would then adjudicate with clearer framing.

This is the governance version of "two people check the math," except one of them is a machine trained to be annoyingly thorough.

The big risk: AI does not remove power, it relocates it

Replacing token voting with AI recommendations can easily become a shell game. Power moves from whales to:

• whoever defines the constitution,
• whoever selects the models,
• whoever controls the evaluation harness,
• whoever controls the data pipelines and access.

If a DAO adopts AI stewards, the real governance question becomes: who governs the stewards?

That implies a need for process around:

• Model provenance: what system is being used, and what are its constraints?
• Reproducibility: can the same prompt and inputs produce comparable outputs?
• Auditability: can the DAO test for systematic bias, manipulation, or silent failures?
• Fallback paths: what happens when the steward is wrong, unavailable, or contested?

Without this, "AI governance" risks becoming governance theater with nicer formatting.

Why this idea is showing up now

Two trends make Buterin's timing unsurprising.

First, DAOs are struggling with legitimacy at scale. Many treasuries are large, many communities are global, and the old pitch of "everyone votes" looks less credible each cycle. Delegation helped, but it also created mini political classes.

Second, LLMs made summarization and structured critique cheap. Tools that were science projects in 2020 are now default workplace utilities. Governance, a domain drowning in text, is an obvious target.

Buterin's move is basically: if the reality is already "a small number of people interpret everything," then at least make that interpretation layer more transparent, testable, and scalable.

Takeaways for builders and token holders

• Token-weighted voting is not broken by accident. It concentrates power because that is what wealth-weighted systems do.
• AI stewards are a coordination tool. Their best case is lowering participation costs and improving decision quality, not magically democratizing power.
• The constitution matters more than the model. If objectives are vague or contradictory, the steward will just produce confident noise.
• Capture risk shifts, it does not disappear. Governance security expands from smart contracts to model selection, evaluation, and information integrity.

What to watch next (practical, not inspirational)

1. DAO pilots with narrow scope: Expect early "steward" deployments on low-stakes domains like proposal summarization, grant rubric scoring, and budget consistency checks. High-stakes treasury control should be a nonstarter, for now.

2. Open evaluation frameworks: The credible path is public test suites, reproducible prompts, and measurable error rates. If a DAO cannot explain how it evaluates steward performance, it is buying vibes.

3. Identity and anti Sybil integration: If a DAO wants to reduce token dominance, it will eventually collide with proof-of-personhood and Sybil resistance (preventing one actor from masquerading as many). AI does not solve that.

4. Governance attack surfaces: Watch for new forms of manipulation, including proposal text engineered to mislead summarizers, data poisoning of steward context, and "model lobbying" where factions fight over which steward is "official."

Buterin's proposal is not a cure for DAO politics. It is an admission that governance is mostly an information problem layered on top of a power problem. AI stewards might help with the first one. The second one still belongs to humans, whether they like it or not.