Subscribe to Bankless or sign in
Ethereum creator Vitalik Buterin was thinking about autonomous agents and DAOs back in 2014, when both were still a distant dream for most everyone in the world.
In his early vision, described in his post "DAOs, DACs, DAs and More: An Incomplete Terminology Guide," DAOs were decentralized entities with “automation at the center and humans at the edges”—organizations that would rely on code, rather than hierarchies of humans, to maintain efficiency and transparency.
Fast forward a decade later, and Variant's Jesse Walden just published "DAOs 2.0," reflecting on how DAOs evolved in practice since Vitalik's early writings here.
In short, Walden noted that the initial wave of DAOs often resembled co-ops, i.e. digital organizations with humans at the center that didn't heavily emphasize automation.
That said, Walden went on to argue that new advancements in AI—particularly large language models (LLMs) and generative models—are now poised to better realize the sorts of decentralized autonomy that Vitalik foresaw 10 years ago.
Yet as DAO experiments increasingly embrace AI agents, we'll face new implications and questions here. Below, let's walk through five key areas that DAOs will have to contend with as they incorporate AI into their approaches.
⚖️ Shifting governance
In Vitalik’s original framing, DAOs were meant to reduce reliance on hierarchical human decision-making by encoding governance rules onchain.
Initially, humans were still “at the edges,” but remained critical for complex judgments. In the DAOs 2.0 world described by Walden, humans still hover at the periphery—providing capital and strategic direction—but the central seat of power increasingly won't be human at all.
This dynamic is poised to redefine governance for many DAOs. We'll still see coalitions of humans negotiating and voting on outcomes, but various operational decisions will increasingly be steered by the learning patterns of AI models. How this balance will be navigated is an open question and design space for now.
🤖 Minimizing model misalignment
Early DAO visions aimed to counteract human biases, corruption, and inefficiency with transparent, immutable code.
Now, a critical challenge is the shift from unreliable human decision-making to ensuring that AI agents are “aligned” with a DAO’s goals. A main vulnerability here is no longer human collusion but rather model misalignment: the risk that an AI-driven DAO optimizes for metrics or behaviors that deviate from human-intended outcomes.
In the DAOs 2.0 paradigm, this problem of alignment—originally philosophical in AI safety circles—becomes a practical issue in economic and governance terms.
This may not be a front-and-center issue for today’s DAOs experimenting with basic AI tooling, but expect it to emerge as a major area of scrutiny and refinement as AI models become more advanced and deeply integrated into decentralized governance structures.
⚔️ New attack surfaces
Consider the recent Freysa contest, where the human p0pular.eth tricked the AI agent Freysa into misinterpreting its “approveTransfer” function to win a $47,000 prize in ether.
Despite Freysa’s built-in safeguard—an explicit instruction never to send the prize—human creativity ultimately outsmarted the model, exploiting the interplay between prompts and code logic until the AI released the funds.
This early contest example underscores how, as DAOs incorporate more complex AI models, they will inherit new attack surfaces as well. Just as Vitalik worried about DOs or DAOs being gamed by colluding humans, now DAOs 2.0 must consider adversarial inputs into the AI’s training data or prompt-engineering attacks.
Manipulating an LLM’s reasoning process, feeding it misleading onchain data, or subtly influencing its parameters could become the new form of “governance takeover," where the battlefield shifts from human-majority voting attacks to more subtle and intricate forms of AI exploitation.
🎯 New centralization questions
The DAOs 2.0 evolution shifts non-trivial power to those who create, train, and control a particular DAO's underlying AI model or models, and this dynamic could lead to new forms of centralized chokepoints.
Of course, training and maintaining advanced AI models requires specialized expertise and infra, so in some orgs to come, we'll see where the direction is ostensibly in the hands of the community but practically in the hands of its skilled specialists.
This is understandable. But going forward it will be interesting to track how DAOs experimenting with AI contend with things like model updates, parameter tuning, and hardware provisioning in this context.
♟️ Strategic vs. operational roles and community buy-in
Walden’s "strategy vs. operations" distinction suggests a long-term equilibrium: AI can handle day-to-day DAO tasks, while humans will provide strategic direction.
However, as AI models become more advanced, they might gradually encroach on a DAO's strategic layer, too. Over time, the “human at the edges” role could shrink even further.
This raises the question: what happens in the next wave of AI-driven DAOs, where in many cases humans might merely supply capital and watch from the sidelines?
In this paradigm, will humans largely become interchangeable investors with minimal influence, shifting away from the co-owned brand approach to something more akin to autonomous economic machines managed by AI?
I think we will see more of a trend in the DAO scene toward org models where humans just play the role of passive shareholders rather than active stewards. Yet with fewer meaningful decisions for humans and the ease of providing capital onchain elsewhere, maintaining community buy-in over time may become an ongoing challenge.
✊ How DAOs can stay proactive
The good news is that all of the challenges discussed above can be tackled proactively. For example:
- On governance — DAOs can experiment with governance mechanisms that reserve certain high-impact decisions for human voters or rotating councils of human experts.
- On misalignment — By treating alignment checks as a recurring operational expense, like security audits, DAOs can ensure an AI agent’s loyalty to communal goals is not a one-off problem but an ongoing responsibility.
- On centralization — DAOs can invest in broader skill-building among its community members. Over time, this would mitigate the risk of governance capture by a handful of “AI wizards” and foster a decentralized approach to technical stewardship.
- On buy-in — As humans become passive stakeholders in more DAOs, these orgs can double down on storytelling, shared missions, and community rituals to transcend the immediate logic of capital allocation and maintain long-term buy-in.
Whatever happens next, it's clear the future here is wide open.
Consider how Vitalik recently introduced Deep Funding, which isn't a DAO effort but is designed to use AI and human judges to pioneer a new funding mechanism for Ethereum open-source development.
This is just one new experiment, but it highlights a broader trend: the intersection of AI and decentralized collaboration is revving up. And as new mechanisms arrive and mature here, we can expect DAOs to increasingly adapt and extend these AI ideas. These innovations will bring unique challenges, so the time to start preparing is now.
