Technical Summary: Voting Biases in Decentralized Autonomous Organization (DAO) Governance

Problem Statement

Decentralized Autonomous Organizations (DAOs) utilize token-weighted voting to allocate resources and set protocol rules, promising efficient and democratic coordination. However, empirical observations indicate that support in DAO voting is strikingly concentrated, with a high prevalence of one-sided outcomes and approval rates often exceeding 75–90%. While prior research has documented frictions such as token concentration and low participation, it remains unclear whether these skewed outcomes stem from substantive community consensus or from systematic biases embedded in the voting interface and proposal presentation. Specifically, it is unknown how much of the observed support is driven by the position of a choice in a list, the stance (approval vs. rejection) of the choice, or the selection of the choice by the proposal author.

Methodology

The study analyzes a cleaned dataset of approximately 51 million votes cast on 127,281 proposals across 19,201 DAO spaces on the Snapshot platform between October 2020 and November 2023.

Data Processing and Stance Classification

• Cleaning: The authors removed invalid, pending, or flagged proposals and votes, filtering for "mature" spaces (those with sufficient followers, voters, and proposal history).
• Vote Expansion: To handle multi-preference voting systems (e.g., weighted, quadratic, approval), the authors mapped ballots to choice-level rows using two schemes: favorite-choice (assigning mass to the top-weighted choice) and support-fraction (dividing mass proportionally). The primary analysis relies on the favorite-choice scheme.
• Stance Detection: A novel pipeline classifies each choice into four stances: Approve, Reject, Abstain, or Other. This was achieved via:
  1. Heuristics: A rule-based keyword classifier with manual review for negations and emojis.
  2. LLM Validation: Large Language Models (Llama 3.1/3.3, GPT-4o) were used to validate the heuristics, showing substantial agreement (Cohen's Kappa ~0.68–0.69).
  3. Human Evaluation: A stratified sample of 590 proposals was manually labeled by 375 crowdworkers, yielding a Kappa of 0.71 with the heuristic approach.

Analytical Approach

The study employs a descriptive and econometric framework to quantify three potential biases:

1. Position Bias: The association between list order and voting-power share.
2. Approval Bias: The systematic advantage of approval-oriented choices.
3. Author Bias: The advantage of choices selected by the proposal author.

The authors use fractional-logit regression models with a quasi-binomial link to estimate the share of voting power for each choice. The dependent variable is the normalized voting-power share (bounded between 0 and 1). The models control for choice rank (linear and squared), stance, author selection, author visibility, and DAO space fixed effects. Standard errors are clustered by DAO space.

To ensure robustness, the authors:

• Subtract the author's own voting power from the total to determine if the "author bias" is merely a mechanical artifact of authors voting for themselves.
• Test various sample specifications, including restricting to top TVL (Total Value Locked) DeFi protocols, excluding low-participation proposals, and varying voting system types.

Key Results

1. Descriptive Findings

• Position Bias: The first-listed choice captures a disproportionate share of voting power (mean share of 70% across all proposals, peaking at 79% for Basic voting systems).
• Approval Bias: Approval-oriented choices overwhelmingly capture the largest share of voting power (approx. 80% overall), but this is heavily correlated with their tendency to be placed in the first position.
• Author Bias: Choices selected by the proposal author receive the largest amount of voting power across all positions and proposal types.

2. Econometric Findings (Average Marginal Effects)

Using regression models to disentangle these overlapping factors, the study finds:

• Author Selection: An author-selected choice gains an average of 58.8 percentage points in voting-power share relative to a non-author choice, holding other factors constant. This is the strongest predictor among the three biases.
• Approval Stance: An approval-oriented choice gains a 27.1 percentage point advantage over non-approval choices.
• List Position: Being listed first (vs. second) yields a 7.7 percentage point advantage.

3. Robustness and Causality

• Author Voting Power: Even after subtracting the author's own voting power from the outcome, the author-choice association remains significant (dropping from ~47.9 to ~32.6 percentage points), indicating the bias is not solely due to authors voting for their own options but likely involves social influence, signaling, or strategic placement.
• TVL Restriction: When restricting the sample to the top 100 DeFi protocols by TVL, the magnitude of the author bias decreases, becoming smaller than the approval bias (though not statistically significantly so), suggesting that in highly liquid/active DAOs, other factors may play a larger role.
• Descriptive Nature: The authors explicitly state these are systematic associations, not proven causal distortions. The results do not prove manipulation but highlight that interface design and author signals are central to how voting power accumulates.

Contributions

1. Empirical Decomposition: This is the first large-scale study to empirically decompose the associations between voting-power share and author selection, list position, and approval stance in DAO governance.
2. Novel Classification: The paper introduces a robust stance classification procedure for DAO choices, triangulated with LLMs and human evaluation, allowing for the systematic analysis of "approval" vs. "reject" dynamics.
3. Design Implications: The study shifts the focus from proposal outcomes alone to the interface and social signals (ordering, author selection, vote visibility) that mediate collective decisions.

Significance and Claims

The paper claims that the high approval rates and one-sided outcomes in DAOs are not merely reflections of community consensus but are systematically associated with the presentation of choices and the identity of the proposer.

• Institutional Design: The authors argue that ordering, author signals, and vote visibility should be treated as institutional design choices rather than neutral implementation details.
• Transparency: The findings suggest a need for clearer information regarding who created a proposal, which option the author selected, and the extent of the author's own voting power to ensure accountability.
• Legitimacy: Systematic advantages for specific choices may penalize dissenting positions and mask limited deliberation, potentially undermining the legitimacy of DAO governance.
• Limitations: The authors maintain a modest stance, acknowledging that the analysis is observational. They note that high approval rates may also reflect pre-vote screening (proposals that fail informal checks never reach the ballot) and that the residual author bias could stem from information cascades, coordination, or unmeasured proposal quality.

In conclusion, the paper provides evidence that DAO voting outcomes are shaped not only by token-weighted aggregation but also by the structural features of the ballot and the social signals emitted by proposal authors.