Voting Biases in Decentralized Autonomous Organization (DAO) Governance

Researchers have found that how votes are presented in decentralized organizations (DAOs) heavily influences outcomes. Specifically, choices listed first, choices that favor the proposal, and choices picked by the person who wrote the proposal get much more support than they otherwise would.

In the world of blockchain governance, the assumption is that token-weighted voting creates a transparent, democratic mechanism. This process allocates millions of dollars in treasury funds. The prevailing belief is that if a proposal passes with 90% support, it represents a genuine consensus. However, this ignores the "black box" of the voting interface itself. We have long understood that token ownership is concentrated. We have not looked closely at whether the way a ballot is rendered drives that concentration.

The illusion of organic consensus

Current DAO governance models assume that once a proposal reaches a formal vote on a platform like Snapshot, the resulting tally is a clean aggregation of preferences. This assumes the interface is a neutral pipe. In reality, the transition from off-chain discussion to an on-chain vote is fraught with invisible filters. Previous studies have noted high approval rates in protocols like Compound or DAOhaus [27, 47]. They largely attribute this to "pre-screening." This is the idea that bad proposals are killed in forums before they ever hit a ballot.

The gap in our understanding is whether the high support we see is a product of substantive agreement or a byproduct of the UI (user interface). As shown in, support is not distributed uniformly.

It is heavily skewed toward specific types of choices. If we only look at the final outcome, we miss the structural mechanics that push voters toward a single conclusion.

Disentangling the three layers of bias

To move past mere observation, the authors developed a methodology to isolate three specific drivers of voting power: position, stance, and authorship. The core of their approach is a fractional-logit model. This model treats the voting-power share of a choice as a bounded outcome between zero and one. This avoids the information loss inherent in turning a weighted vote into a binary "pass/fail" metric.

The researchers built a sophisticated choice stance detection pipeline to solve the problem of unstructured data. Since choices can be anything from a "Yes" string to a "👍" emoji, they used a three-stage architecture: 1. A registry-backed heuristic engine. This uses multi-language keyword matching (exact, prefix, suffix, and high-confidence containment). 2. Validation via Large Language Models (LLMs). They used Llama 3.1, Llama 3.3, and GPT-4o to ensure the heuristics were accurate. 3. Ground-truth verification using human crowdworkers via Prolific. This confirmed the classification accuracy.

By combining these, they achieved a Cohen’s Kappa of 0.71 between heuristics and humans. This metric indicates substantial agreement. This allowed them to map complex, multi-preference ballots into a standardized format. They could then mathematically decompose the influence of each bias.

Author signals dominate the ballot

The results are striking. The authors report that the strongest predictor of a choice receiving high voting power is whether the proposal author selected it. This "author bias" yields a 58.8 percentage point increase in voting-power share. This means an author-selected choice gains nearly 60% more voting power than a similar choice not chosen by the author.

Even when the researchers performed a sensitivity analysis, the association remained robust. They subtracted the author's own voting power from the total. This tests if the effect was just the author "buying" their own choice. The author-choice advantage dropped from 47.92 to 32.56 percentage points. However, the effect did not vanish [Table 6]. This suggests the signal involves social influence or reputational cues.

Following author signals, the "approval bias" provides a 27.1% advantage. This is the tendency to favor choices that support the proposal. Finally, "position bias" accounts for a 7.7% bump .

This is the simple advantage of being listed first in the UI. As seen in [Figure 4B], the author boost is remarkably stable across different positions and stances. It is the primary driver of skewed outcomes.

Limits of the observational lens

There are two major caveats to consider. First, this is an observational study, not a controlled experiment. The authors state they are measuring systematic associations. They are not proving causal distortion. We cannot definitively say a voter changed their mind because a choice was listed first. We can only say first-listed choices receive more votes.

Second, the "survival bias" problem remains. The dataset only includes proposals that reached the Snapshot voting stage. We only see the survivors of an off-chain filtering process. High approval rates may reflect proposal screening. This is where controversial proposals are abandoned in Discord or forums before a formal ballot. This means the baseline of DAO governance might be more contentious than the Snapshot data suggests.

Verdict: Treat the UI as an institutional layer

The evidence suggests the "wisdom of the crowd" in DAOs is heavily mediated by the layout of the ballot. If you build a decentralized system, you cannot ignore the social signaling of the author. You also cannot ignore the psychological weight of list order.

For practitioners, this means moving toward more auditable interfaces. This could include shuffling choice order or anonymizing author selections. It could also mean providing clear disclosures. Users should know how much of a winner's margin comes from the author's own weight. Until then, a 90% approval rate should be treated with skepticism. It may not be a mandate, but a reflection of the interface.

Figures from the paper