DAO votes strongly favor proposal authors’ chosen options, study finds

Decentralized autonomous organizations, or DAOs, distribute decision-making through governance tokens rather than conventional managerial authority. But whoever writes a proposal still chooses which alternatives appear on the ballot and how they are presented. A new arXiv preprint by Stefano Balietti, Pietro Saggese and Markus Strohmaier examines how that role relates to voting outcomes.

Their central finding is a strong association between an option being selected by the proposal’s author and its eventual share of voting power. That association is larger than the corresponding associations with approving the proposal or appearing first on the ballot. Crucially, the researchers use “bias” to describe statistical patterns, not proof that voters’ preferences were causally distorted.

Looking inside the ballot

The cleaned dataset contains 127,281 proposals and approximately 50.8 million votes across 19,201 Snapshot spaces, covering October 2020 through November 2023. Snapshot records votes off-chain while using blockchain addresses and governance rules to establish voting power. In its default configuration, voters can see the author’s identity, individual votes and running totals.

The researchers examine individual ballot options rather than simply whether proposals pass. They identify three features: an option’s position, whether it approves the proposal, and whether the address that created the proposal voted for it. “Author-selected” therefore refers to a recorded vote, not an inferred preference from the proposal’s wording.

To classify options as approving, rejecting, abstaining or other, the team combines multilingual rules with contextual review. Language-model outputs and human annotations provide validation checks, with substantial but imperfect agreement between the rules and human labels.

The main regression analysis covers single-preference voting systems, which account for 93% of proposals. Its fractional-logit models accommodate voting-power shares between zero and one. Controls include ballot length, option-text length, voting system and whether the author participated. DAO-level fixed effects account for persistent differences between communities, while uncertainty estimates allow observations within a community to be related.

The largest association belongs to authors

In the main specification, the reported average marginal association for author selection is 58.8 percentage points of voting-power share. This comparison is averaged over options in proposals where the author voted, contrasting author-selected with non-author-selected options while holding other modeled characteristics fixed.

The corresponding estimate for approving options is 27.1 percentage points. For position, the estimate is 7.7 percentage points for appearing first rather than second—not first rather than every other position combined. These are model-based comparisons, not experimentally measured gains from changing a ballot.

Part of the author association is mechanical: authors contribute their own voting power to their chosen options. The researchers therefore conduct a separate comparison using proposals where the author voted and other voters contributed positive voting power.

Within that diagnostic sample, the author-selection estimate falls from 45.6 to 29.7 percentage points when shares are recalculated using only non-author votes. The reduction is roughly one third of the original estimate. The remaining association reflects support from other voters, but does not establish that the author caused that support.

The pattern also varies across settings. Excluding proposals with fewer than two votes reduces the author estimate to 39.7 percentage points. In subsets matched to large decentralized-finance protocols ranked by historical Ethereum total value locked, author estimates range from about 27 to 33 points. Approval estimates are larger in those subsets, although the paper cautions that the difference between the author and approval associations is not statistically significant there.

Strong support is not necessarily independent consensus

Several explanations remain compatible with the evidence. Voters might follow a visible author vote, defer to relevant expertise, coordinate beforehand or independently favor a high-quality option. Authors may also anticipate community preferences. Their tendency to vote early limits explanations based solely on following already-recorded votes, but does not distinguish these other mechanisms.

Selection before voting creates another boundary. Proposals may be discussed, revised or abandoned before reaching Snapshot. The dataset observes formal ballots, not all ideas considered by a community, so high approval can partly reflect earlier screening.

The authors suggest explicitly reporting proposers’ voting power, randomizing non-default option order and hiding interim tallies where herding is a concern. These are proposed design responses, not interventions tested by this study. The preprint’s narrower conclusion is that author participation, ballot structure and visible voting activity deserve scrutiny before lopsided outcomes are interpreted as independent consensus.