The Power to Change the Rules Is Also for Sale — When Governance Itself Becomes the Attack Surface

If Code Is the Rules, What Is the Table That Changes the Rules?

When we analyzed over 370 proposals to change the rules and millions of records left in the ledger across two major decentralized finance (DeFi) services, a group of roughly 3 to 5 voters was enough to sway the direction of a majority of the proposals (arXiv:2305.17655, preprint). Decentralized finance is financial services like lending and trading that run on smart contracts. The takeaway from this number for this article is simple: at least in these two places, the party trying to buy votes had to deal with fewer people than expected.

A DAO (Decentralized Autonomous Organization) is an organization without a boss or a board of directors that writes its rules into smart contracts and changes those rules through votes by participants holding tokens. Smart contracts, as defined in "What Is a Smart Contract?", are conditions and actions to be taken under those conditions written into a ledger in advance, which are executed exactly as written without human judgment when someone calls them. The investment club The DAO placed at the beginning of that article was also an organization that used votes to decide on investment targets. That article noted that in a contract written to be modifiable, the person holding the key to make modifications becomes a new seat of trust, whereas a DAO is an organization that entrusts that key for changing rules to a vote instead of a single person. Governance is the system of who sets and changes rules and through what procedures, and in a DAO, that procedure is precisely voting.

The treasury is the funds jointly owned by the DAO, and when a proposal passes, money is paid out from there. By the time research systematically organizing DAO attacks came out, DAOs had accumulated large assets in their treasuries and were actually operating several widely used decentralized finance services (arXiv:2406.15071, preprint). A governance token is a token that grants votes in the organization's balloting proportional to how many are held, and participants pass or reject proposals with those votes. Who participates and how, and how much influence each person holds, are specified in the smart contract that records the organization's voting rules—namely, the governance contract (arXiv:2604.25959). In a DAO that ballots on the ledger, proposals, votes, and funds leaving the treasury are all recorded indelibly on the ledger (DOI: 10.3389/fbloc.2026.1890169). However, there are also DAOs that collect ballots using tools outside the ledger (DOI: 10.2139/ssrn.7103878, preprint).

The point that holding tokens grants an equivalent amount of votes resembles proof of stake from "The Right to Write a Line in the Ledger", which is a method of reaching consensus by proving one's own share held in the ledger instead of through computation. However, what the share in proof of stake determines is whose turn it is to write the next page of the ledger, whereas what the votes of a governance token determine are the rules of the organization.

A natural expectation arises here. Because the rules are in the code and the code is changed only by voting, the expectation is that a small minority cannot change them at will. The question this article follows is the exact opposite. If voting rights themselves are items that can be bought and sold in the market, can the authority to change the rules be bought as well?

This article does not evaluate any organizations or tokens, nor does it write out attack procedures; the vote-buying and dark DAOs that appear later go only as far as the fact that such structures are academically formalized and their implications. All figures are values observed by that research from that data at that point in time, and do not represent the state of any organization today.

Structures where administrative authority left in the code serves directly as an exit are covered in "A Rug Pull Is Not a Hack". What we look at here is the procedure for changing rules itself. Even without administrator keys or backdoors left behind secretly, this is a problem that holds true simply if voting operates normally. The authorities left in the contract can be extracted as a list by reading the code. Voting authority is also written in the code, but why the person holding those votes casts them that way is not written in the code.

When we gather and categorize actual attacks that occurred, attacks proposed in theory, and attacks blocked after being discovered in audits, this contrast becomes starkly apparent. Here, audits refer to code audits from "What Is a Smart Contract?", which are the process of humans or tools reading code to find flaws before uploading the contract to the ledger. A significant portion of attacks targeting DAOs exploit the less-touched, more complex human properties entangled with governance, whereas audits tend to lean toward vulnerabilities in code and protocols (arXiv:2406.15071, preprint). This article departs from this asymmetry.

Votes Are Scattered, But They Do Not Show Up

If we translate the expectation a bit more precisely, it goes like this: if tokens are widely scattered, a single person cannot make a majority, and to make a majority, they would have to buy them up from the market to that extent, but that price is too expensive. Therefore, the rules are safe. One condition is hidden in this logic: the condition that the people holding the votes actually vote.

Actual records run counter to this condition. In the governance of decentralized finance platforms, users rarely used their governance tokens for voting, voter turnout showed a negative correlation with transaction fees, and voting was highly centralized (arXiv:2308.04267, preprint). Because a vote on the ledger is also a transaction that writes a single line to the ledger, it costs the fee from "The Right to Write a Line in the Ledger"—in other words, the price attached to having your one line included on the next page. This means that voter turnout was low when those fees were expensive. The pattern of votes concentrating in the hands of a few people is the same concentration across layers that "Is Blockchain Really Decentralized" measured, appearing here at the governance layer as well.

Votes that are not cast are left out of the majority calculation. Therefore, how much one must buy to become a majority depends not on the entire issuance of tokens, but on the amount of votes that actually show up, and the lower the participation, the lower that price drops.

Low participation does not end as a one-time exception. When sustainability indicators for DAOs were built using ledger records, low participation rates and high proposer concentration appeared as a repeating pattern (arXiv:2504.11341, preprint). Nor does it go up easily just by tweaking the procedures. Across 1,959 proposals from five major DAOs, the length of the voting window—the period during which votes are received—showed no statistically or economically meaningful relationship with participation. What explained most of the variation in participation was long-standing institutional characteristics unique to each DAO (DOI: 10.2139/ssrn.6588294, preprint). This suggests it may be a problem that is difficult to fix simply by turning the knobs of procedure.

Votes Are Transferable—And They Actually Are Transferred

Governance tokens can be passed to others, and when the tokens change hands, the votes go with them. Tracking the movement and holding of tokens reveals a flow where governance tokens are used in a single action and then sold off immediately. There is also evidence that certain platforms used another platform's governance procedures to push through their own agenda at the expense of that platform's interests (arXiv:2308.04267, preprint). If there is no need to hold onto a vote for long, what determines the value of a vote is not the organization's future, but the value created by that single decision.

There are also signs of tokens moving just before a vote. Out of 8,116 evaluated proposals, a tendency for governance token ownership to shift right before a vote was observed in 1,202 cases (14.81%). In the same study, there were DAOs—making up at least 20.41% of the total—where internal contributors who had received tokens while contributing to the organization single-handedly decided at least one proposal (arXiv:2309.14232, preprint).

Votes can also be borrowed rather than bought. A 2020 preprint showed that it was feasible to attack the governance design of the decentralized finance service with the largest market share at the time and take over the entire service. According to the authors' calculations, it was a path that, in principle, required only two transactions, tied up zero assets, allowed an attacker to drain 500 million dollars in collateral, and endlessly minted the service's stablecoin. A stablecoin is a coin designed to peg its value to a standard like the US dollar ("The Code Did Exactly What It Promised"). Assets did not need to be tied up because of flash loans. A flash loan is a loan borrowed without collateral and paid back within the same transaction; if it is not paid back, the entire transaction is undone ("A Smart Contract Cannot See the Outside World"). About two weeks after the authors responsibly disclosed the attack details, the service changed its governance settings to mitigate this attack path (arXiv:2002.08099, preprint). The structural vulnerability was formally identified first, and the design was changed afterward.

Research on the defense side starts from the same problem. If the number of tokens is used as the weight of a vote, it creates an opening to manipulate voting results using flash loans and pass a malicious proposal that drains treasury funds into an attacker's wallet. As a defense against this, a time-weighted snapshot structure was proposed (arXiv:2505.00888, preprint). As the name suggests, it is a structure that reflects holding over time into the vote's weight instead of a single balance at the moment of the vote.

The idea of separating voting rights from economic interests is older than blockchains. Empty voting is a state where one holds the votes but does not bear the gains and losses produced by that decision, a term coined in corporate governance. It is created by borrowing shares on the record date for counting voting rights, offsetting gains and losses tied to the stock price with hedges, or selling shares between the record date and the voting day. A 2008 modeling paper (preprint) modeled corporate voting when an informed trader, such as a hedge fund, holds shares and votes separately in this way. The presence of such a trader could actually improve overall efficiency, but the model also included cases where the trader went into a net short state—where they sold more shares than they held so they profited if the stock price fell—and then cast their vote in favor of lowering the corporate value. The authors' conclusion is conditional. If the cost of detaching votes from shares is low and other shareholders are likely to vote the right way, empty voting tends to reduce efficiency. Conversely, if the cost of detachment is high or other shareholders' votes do not align well with the right decision, efficiency can improve (DOI: 10.2139/ssrn.1108632, preprint).

This model deals with corporate voting in the stock market. What we take from it is the structure. The structure formalized in stocks, where votes and economic interests are split, exists in the exact same form in token voting. The flash loan path mentioned earlier was a way to hold votes without tying up assets (arXiv:2002.08099, preprint). Whether the model's efficiency conclusions hold under the same conditions in token voting is outside the scope of what that model examined.

There are also models that deal with token voting itself. This research formulated a theoretical model of the conflict of interest between a single large holder and many small participants under token voting where tokens are bought and sold strategically, and verified its predictions using voting data from major DAOs. The more concentrated ownership was in one place, the worse the effect on platform growth, and there were three conditions that mitigated this effect: the size of the platform, the illiquidity of the tokens, and long-term incentives (DOI: 10.1287/mnsc.2024.07033). Illiquidity is the degree to which tokens are difficult to buy and sell. The fact that the negative effects of concentrated ownership are mitigated the harder it is to buy and sell tokens means, conversely, that the more easily votes can be bought and sold, the more one of the forces mitigating that effect disappears.

Among DAO design proposals, some view empty voting as a distinct threat to be guarded against. One design proposal, which suggests creating temporary and contestable control through sequential auctions, cites robustness against empty voting as one of the advantages of its approach (arXiv:2403.16980, preprint).

Even the Structure That Hides Bribes Is Formalized

Decentralization, as defined in "The Right to Write a Line in the Ledger", is a state where power is not concentrated in one or two places but scattered among multiple independent participants. A new metric has emerged for measuring decentralization in a DAO. Voting-Bloc Entropy (VBE) is a metric that measures how similar participants' utility functions are across various proposals (arXiv:2311.03530, preprint). A utility function represents how much each person desires a particular outcome. The authors' critical concern is that existing definitions of decentralization fail to capture the core properties of diverse and even participation, and this metric looks at how similar interests are instead of wallet counts or token distributions (arXiv:2311.03530, preprint).

Plugging things into this framework measures previously disconnected phenomena with a single ruler. The authors use VBE to prove several results regarding the effects that delegating votes to others, bundling multiple agenda items into a single proposal, and vote-buying have on decentralization. Here, vote-buying refers to the act of purchasing another person's vote in exchange for compensation. Among them, the result emphasized by the authors is this: the greater the decentralization measured by this metric, the greater the risk of systematic vote-buying (arXiv:2311.03530, preprint). This goes in the opposite direction of the intuition that decentralization equals safety.

To show that this threat is realistic, the authors present the first practical implementation of a dark DAO. A dark DAO is a device proposed to invisibly corrupt identity systems, including those used in DAO voting. The prototype targets a DAO on Ethereum by using a blockchain's Trusted Execution Environment (TEE), an area isolated inside a chip where computation contents cannot be seen from the outside. The authors' conclusion is measured: a dark DAO is something that DAO governance must realistically worry about moving forward. From the same analysis, the authors also derive practical proposals to increase DAO decentralization (arXiv:2311.03530, preprint).

What this article carries over is up to the nature of the results. If transactions transferring votes can leave no trace on the ledger, the ledger cannot answer the question of whether these votes were sold. What is lost at that point is observability—in other words, the means to verify by looking at the ledger.

This tension is written more sharply in a survey reviewing DAO governance protocols (preprint). That survey diagnosed that in the DAO voting methods of the time, buying and selling votes as well as coercion were easy, vote privacy did not exist or was maintained only very briefly, and verifiability was traded off against privacy (arXiv:2406.08605). "If We Voted on a Blockchain, Nobody Could Tamper With It" addressed this same tension in elections. The "no receipt" property in that article is a feature that prevents voters from proving to others how they voted even if they want to, which was necessary because if they could prove it, they could also prove it to a vote buyer. The end-to-end verifiable voting introduced in that article was a method that allowed voters to confirm that their ballot was recorded and included in the tally while hiding the contents of the vote. The DAO voting methods observed by that survey lie on the side that has chosen result verification instead of hiding votes—meaning the votes are public on the ledger. When votes are visible, anyone can verify the outcome, but it is also visible whether the promise to transfer the vote was fulfilled. Token voting has one more path that elections do not. Election votes are tied to people, but governance tokens allow the votes themselves to be moved. Among the four critical attack vectors identified by a study summarizing the security of DAOs across multiple blockchains (preprint), the top item is also vote-buying attacks (arXiv:2607.16548).

The most dryly written conclusion of this section comes from a preprint analyzing the relationship between prediction markets and decentralized oracle governance. Prediction markets are markets where people bet money on the outcomes of future events. Oracles are entities that feed off-chain facts into smart contracts, as defined in "What Is a Smart Contract?". In this study, token voting determined the outcomes of 134 disputed contracts, and the author analyzed 4,992 of those voting records. By tracking addresses, 11% of the voters could be linked to trading addresses, and among them, not a single instance was found where the same person held a stake wagered on a market at the very same time they cast a vote in that market. Instead, the study also reports what it found alongside this: the top 3 individuals held 51% of the total voting weight, and the top voters formed a single voting block with highly correlated behavior (average pairwise correlation $\varphi \approx 0.86$). Translated into the language of VBE, a cluster of aligned interests was actually observed. The author's conclusions are twofold. One is that when there is no objective correct answer, prediction market prices act as a prominent focal point that oracle voting relies on together. The other is that this does not mean the judgment is clean; rather, the ledger alone cannot make that determination. Governance capture is a situation where a minority seizes voting to suit their own interests. Individual-level governance capture cannot be structurally audited using public ledger data alone (DOI: 10.2139/ssrn.6874201, preprint).

If ledger data alone cannot distinguish between something not existing and something not being visible, the list of questions the ledger can answer is shorter than expected.

What Prevents What

Now let's place each defense mechanism on the scale to see what each prevents and what it cannot stop.

A DAO quorum is the rule that a vote must receive at least this many votes to be valid. If the quorum in "Consensus Without Trust" was the predetermined number of participants sufficient to finalize a decision, here it is counted in tokens instead of headcounts. When analyzing the governance design on the ledger of 204 DAOs through signaling theory—which reads design choices as signals sent to the outside—having a minimum token quorum showed a positive relationship with token performance. Attaching voting rewards such as NFTs or additional governance tokens showed a negative one (DOI: 10.1177/02683962251343627). What a quorum prevents is being quietly passed when nobody is looking. It cannot prevent the majority that satisfies the quorum itself, and in organizations with low participation, raising the quorum can increase the number of proposals that fail to pass, potentially bringing the organization to a halt.

A timelock is a mechanism that forcibly leaves a certain amount of time empty between a vote passing and its actual execution. An earlier study proposing a security framework for DAOs spanning multiple blockchains includes a mechanism of the same grain in its recovery layer—specifically, a mechanism allowing decisions to be reversed within a set time, along with staged dispute resolution (arXiv:2607.16548, preprint). Time-weighted snapshots introduced as a defense against flash loans are on the same side in that they use time for defense. The study introducing that defense finds the weakness of existing flash loan defenses in accessibility and argues that this weakness harms security (arXiv:2505.00888, preprint). What a timelock prevents is immediate execution, and what it cannot stop is the majority itself. A mechanism that buys time only works when someone is actually doing something during that time.

Delegation is a system where token holders hand over their voting power to someone else to vote on their behalf. A delegate is a person who receives delegation and votes on behalf of others. Examining the design and implementation of 48 public DAOs deployed on Ethereum, three mechanisms introduced to enhance security or participation—namely, token registration, staking, and delegation—contribute to the concentration of voting power. The authors write that this is not a problem of a specific DAO, but that governance mechanisms themselves systematically reinforce centralization (arXiv:2604.25959, preprint). Here, staking has the same name as the staking in "The Right to Write a Line in the Ledger"—meaning depositing a stake in the ledger to earn the right to write the next page—but it is not a right to write the ledger, rather a mechanism hooked onto that organization's governance contract.

The problem arises when the delegate's interests diverge from those of the organization. In a preprint analyzing the delegation design of a decentralized finance organization, delegates showed expertise, participated more in votes, and made more informed choices. However, when their personal interests diverged sufficiently from the organization's interests and they held enough votes to sway the outcome, they could prioritize personal gain over organizational gain. The market generally rewarded correct votes and punished incorrect ones by newly granting or withdrawing delegation. Still, in the authors' phrasing, even with holdings fully transparent, consistently punishing a delegate whose interests diverge remained difficult (DOI: 10.2139/ssrn.4533512, preprint).

Monitoring capacity is also a bottleneck. Decentralized governance is monitoring-heavy, so if the speed at which proposals flood in exceeds the capacity that broad participation can keep up with, practical control can flow toward a very active minority. In quarterly data by DAO, once proposal activity crossed a certain threshold, the response of active voters slowed down statistically significantly, and the author views this result as aligning with the mechanism of "too big to monitor" (arXiv:2603.11222, preprint). This means that the more active an organization becomes in churning out proposals, the narrower practical control can become. When practical control narrows, those attempting to buy votes also have fewer people to deal with.

The treasury is where voting most directly translates into money. A study analyzing blockchain treasury governance through comparative institutional economics frames this as a balance between the cost of dictatorship and the cost of disorder. Entrusting it to one side risks abuse; entrusting it to no one makes chaos the cost. Those costs shift along the community's lifespan, typically surfacing or being learned through governance crises (DOI: 10.1111/rego.12659). Read differently, these costs are hard to see before an accident happens. There is also a record of recalculating whether the outcome changes if vote weights are counted differently. Selecting 15 high-participation proposals from a major DAO and recalculating by pressing each wallet's vote weight down to its square root flipped the winners in 4 cases (26.7%). All four flipped cases were resource allocation decisions such as funding, grants, and security expenditures (DOI: 10.2139/ssrn.7103878, preprint). This is a property observed in a small sample of 15 high-participation proposals, and is not a normative claim about which calculation is correct.

Meta-governance is a structure where one DAO holds another DAO's governance tokens and intervenes in that DAO's voting. Starting from an initial 16 DAOs on Ethereum and tracing out these relationships, a network consisting of 61 DAOs and 72 meta-governance relationships emerged. The place where decisions are actually made can shift elsewhere, and the composition of the people actually casting votes can become blurred. Meta-governance obscures the context of voting and draws in actors who move according to their own interests (arXiv:2603.00708, preprint). All the votes this study looked at are on the ledger. However, if the address that cast that vote is another organization, and that organization's vote is swayed by yet another organization, the question of whose intent this vote represents does not end with a ledger lookup.

The simplest yet most uncomfortable result lies in the proposal documents. Running automated detection across 16,427 DAOs and 122,307 proposals spanning 9 blockchains, over 60% of the reviewed proposals failed to provide members with matching explanations and code together (arXiv:2403.11758, preprint). Even if votes are well-distributed, participation is high, quorums are generous, and timelocks are long, it is difficult to call the sum of those votes consent if voters do not know what the code they are passing actually does. The legitimacy of a procedure relies on knowing what is being passed, but in those proposals, that knowledge is severed right from the documentation.

The alternative most frequently proposed is quadratic voting. As defined in "Is Blockchain Really Decentralized", quadratic voting is a rule that dampens the influence of large holders by making votes grow less as holdings increase. Strong negatives have emerged as to whether this rule actually raises attack costs. A 2026 preprint proves an impossibility result stating that on blockchains where anyone can create wallets without permission, no rule that draws voting power solely from wallet balances can succeed. A Sybil attack is an attack where fake identities are created in bulk to inflate headcounts, as defined in "What Is a Blockchain", and here it means splitting tokens across multiple wallets. As long as even a single wallet receives votes, no matter how gently one chooses the vote-dampening function, the splitting party's total voting power grows at least proportionally with their holdings (arXiv:2605.18990, preprint). Choosing a vote-dampening function cannot solve this problem.

Therefore, alternative designs bring in a second tool from outside the balance. One of them is a time constraint, a design that requires tokens to be locked up for a certain period to cast a vote. The idea is that when the opportunity cost for large holders to lock up their tokens is high, it operates to suppress the power of large holders. Since rules looking only at a single balance will not do, something from outside the balance—whether time, identity, or reputation—must be added. This impossibility result does not answer whether that something actually defeats the token-splitting party, and identity and reputation cannot be proven solely through the ledger the way balances are.

How Concentration Is Measured and Where Failures Happen

The very first thing that shakes is measurement. Voting rights are widely distributed, but a minority actually exercises that power; one preprint calls this "shadow centralization" and argues that the concentration of holding and the concentration of exercising must be measured separately. As concentration increased in the transition from holding to exercising, the performance of the applications operated by that DAO grew worse. That effect came with conditions. When voters were more experienced and better connected to one another than token holders as a whole, the bad effect was mitigated; it was more pronounced in proposals requiring extensive deliberation and weighing of pros and cons, and less pronounced in novel types of proposals (DOI: 10.2139/ssrn.6162109, preprint). VBE uses an entirely different axis. Instead of measuring how scattered power is, it measures how similar interests are (arXiv:2311.03530, preprint). Change the indicator, and the evaluation of the same organization changes.

The place where discussions actually happen is not just on the ledger either. Discussions mostly take place in forums and tools outside the ledger, and there are recorded cracks in that layer too. In the forum of a DAO operating a large treasury, the property where forum activity predicted whether a vote would be tightly contested disappeared after a specific point in time. The authors view this change as coinciding with the emergence of language models that write text indistinguishable from human writing. And they use a model to show that the signal value of a forum collapses discontinuously at a certain threshold and that the collapse does not reverse. They also note that agents relying on AI hold vastly fewer voting rights, which does not fit the hypothesis of insiders seizing power (DOI: 10.2139/ssrn.7284923, preprint). Since it is a single preprint and a sample of one organization, it is safer to read it for its structural implications rather than concrete figures. The deliberation layer can break down separately from the voting layer.

As a result after alternative designs are actually deployed, there is convergence. When comparing three large DAOs through a qualitative case study, the three organizations had independently converged on the same three-tier structure: a legal foundation, a security committee, and token-holder governance (DOI: 10.3389/fbloc.2026.1890169). The legal foundation and the security committee are institutions outside the code. It is a direction where organizations that claimed they would run purely on code are bringing such institutions back in.

There are also records where participation did not drop. A study analyzing over 3,000 proposals over 20 months within another blockchain's governance framework (preprint) reports that, unlike studies reporting that participation dropped over time in other frameworks, participation here was maintained or even increased (arXiv:2507.20234). This means that declining participation is not an inevitability of DAOs, but can vary depending on design.

Whether the failure lies in the design or in the people is also divided. One preprint rejects the common explanation that DAO governance failures are due to flawed mechanisms or insufficient incentives. Instead, failure follows predictably from a systematic mismatch between the profile of participants embedded in governance mechanisms and the cognitive and social structures of real humans. The authors organize dimensions of failure using secondary evidence from five protocols, and write that formal decentralization produces functional recentralization in all of them (DOI: 10.2139/ssrn.6797979, preprint). Because it is a descriptive analysis relying on secondary evidence, it does not empirically prove that causality. What this analysis leaves behind is the direction of the question: is it a matter of fixing the mechanisms, or does the person assumed by the mechanisms not exist to begin with?

What Can Be Verified and What Cannot Be Verified

There are things that can be verified. In a DAO where voting takes place on a ledger, who holds how much, which address cast a vote on which proposal, when and how, and how much left the treasury and where it went are all recorded on the ledger, and anyone can check this at any time without having to trust anyone else's word. It is rare in traditional organizations for an outsider to receive and analyze such detailed voting records in their entirety, and that is why most of the empirical studies cited in this article were possible.

There are also things that cannot be verified. Those include whose will a vote represents (arXiv:2603.00708), whether there was compensation attached to that vote (arXiv:2311.03530), whether the person who cast the vote understood what they were passing, and whether governance was captured at the individual level. A voter's understanding is not left on the ledger, and over 60% of the reviewed proposals failed to properly match the explanations and code that should form the basis of that understanding (arXiv:2403.11758, preprint). The last of these cannot be structurally audited using ledger data alone (DOI: 10.2139/ssrn.6874201, preprint).

The boundary between the two lists is this: what can be verified is the distribution of votes, and what cannot be verified is the consensus behind those votes.

The phrase "code is law" is true up to the point that the rules are in the code. That does not mean the authority to change the rules lies nowhere outside the code. The procedures for changing the rules are written in the code, but the votes that drive those procedures move around in the market (arXiv:2308.04267, preprint), can be temporarily borrowed without collateral (arXiv:2002.08099, preprint), are handed over under the name of delegation (arXiv:2604.25959, preprint), and can be held in the hands of other organizations (arXiv:2603.00708, preprint). None of those are bugs in the code; they are all features working exactly as designed.

Therefore, the question we must ask is not whether this organization is decentralized, but rather what guarantees what.

Quorums stop things from quietly passing when no one is watching, but they cannot stop a majority itself. Timelocks prevent instant execution, but if no one is watching during that time, they prevent nothing. Delegation is a mechanism introduced to increase participation, but it contributes to new concentration (arXiv:2604.25959, preprint), and it is difficult to consistently punish misaligned agents (DOI: 10.2139/ssrn.4533512, preprint). Quadratic voting dampens the votes of large holders, but as long as it draws power solely from balances, it cannot beat splitting wallets (arXiv:2605.18990, preprint).

Each mechanism guarantees something specific. However, even if you add up all those items on the list, the sentence "this decision represents the will of this community" does not emerge. That sentence is not the kind of sentence that a ledger can prove.

What cryptography and blockchains do is make it possible to verify what procedures were followed for something claimed to be a consensus, which is different from creating the consensus itself. This body of literature continues the work of not stopping at measuring the distribution of votes, but asking where those votes came from.

The statement that the rules are in the code did not mean that the authority to change the rules is trapped inside the code.

Source Papers

• "SoK: Attacks on DAOs" (2024) — arXiv:2406.15071 (preprint) — A classification combining real, theoretical, and audit-blocking attacks. Many attacks target the human layer, whereas audits look at the code layer—an asymmetry that serves as the starting point of this piece.
• "Understanding Blockchain Governance: Analyzing Decentralized Voting to Amend DeFi Smart Contracts" (2023) — arXiv:2305.17655 (preprint) — Over 370 proposals and millions of on-chain events. Just 3 to 5 people can sway the direction of a majority of proposals.
• "The Vulnerable Nature of Decentralized Governance in DeFi" (2023) — arXiv:2308.04267 (preprint) — Low voter turnout, a negative correlation with fees, and centralization. Observation of a "use once and immediately sell" dynamic.
• "On the Centralization of Governance Power in Decentralized Autonomous Organizations" (2026) — arXiv:2604.25959 (preprint) — 48 public DAOs. Token registration, staking, and delegation systematically reinforce centralization.
• "The Governance of Decentralized Autonomous Organizations: A Study of Contributors' Influence, Networks, and Shifts" (2023) — arXiv:2309.14232 (preprint) — Ownership shifts right before voting in 1,202 out of 8,116 proposals (14.81%).
• "The Decentralized Financial Crisis" (2020) — arXiv:2002.08099 (preprint) — Formally specifies that governance takeover is feasible through two transactions utilizing flash loans, followed by public disclosure of responsibility and parameter modification about two weeks later.
• "Balancing Security and Liquidity: A Time-Weighted Snapshot Framework for DAO Governance Voting" (2025) — arXiv:2505.00888 (preprint) — Time-weighted snapshot defense. Traces the weaknesses of existing defenses to accessibility.
• "Empty Voting and the Efficiency of Corporate Governance" (2008) — DOI: 10.2139/ssrn.1108632 (preprint) — Conditional conclusion that efficiency drops when the cost of separating votes from shares is cheap.
• "Decentralized Autonomous Organization Governance" (2026) — DOI: 10.1287/mnsc.2024.07033 — Three conditions that mitigate the negative effects of ownership concentration: platform scale, token illiquidity, and long-term incentives.
• "Economic DAO Governance: A Contestable Control Approach" (2024) — arXiv:2403.16980 (preprint) — A design creating temporary control through sequential auctions. Explicitly sets defending against empty voting robustness as a goal.
• "DAO Decentralization: Voting-Bloc Entropy, Bribery, and Dark DAOs" (2023) — arXiv:2311.03530 (preprint) — The VBE metric. Results showing that greater decentralization increases the risk of organized vote-buying, along with the first practical implementation of dark DAOs and suggested improvements.
• "Perils of current DAO governance" (2024) — arXiv:2406.08605 (preprint) — Diagnosis that vote-buying and coercion are easy, voting privacy is absent or brief, and verifiability is achieved at the expense of privacy.
• "Towards Secure and Trustworthy DAOs for Cross-Chain Governance" (2026) — arXiv:2607.16548 (preprint) — Vote-buying is the first of four deterministic attack vectors. Recovery layers (time-lock reversals, staged dispute resolution).
• "Price as Truth: Market Prices, Oracle Governance, and the Limits of Decentralized Dispute Resolution" (2026) — DOI: 10.2139/ssrn.6874201 (preprint) — 4,992 voting records. Zero instances of simultaneous position-holding and voting, top 3 holders command 51%, and individual-level capture cannot be structurally audited using public ledgers alone.
• "The power of governance: A study on the relationship between on-chain DAO governance and token performance" (2025) — DOI: 10.1177/02683962251343627 — 204 DAOs. Minimum token quorum shows a positive relationship, while voting incentives show a negative relationship.
• "Reintermediation in Decentralized Governance" (2023) — DOI: 10.2139/ssrn.4533512 (preprint) — Voters who delegated were more professional and participated more, but could put personal interests first if interests diverged significantly. Even with transparency, consistent punishment was difficult.
• "Monitoring Limits in DAO Governance: Capacity Breakpoints and Endogenous Concentration" (2026) — arXiv:2603.11222 (preprint) — Monitoring intensity and the "too big to monitor" mechanism.
• "The Blockchain Treasury Governance Dilemma" (2025) — DOI: 10.1111/rego.12659 — Viewing treasury governance as a balance between the costs of dictatorship and the costs of anarchy. Costs are generally learned through crises.
• "An Empirical Evaluation of Quadratic Voting Counterfactuals in Token-weighted Governance" (2026) — DOI: 10.2139/ssrn.7103878 (preprint) — Winner reversals in 4 out of 15 proposals (26.7%), with all four being resource allocation agenda items.
• "The On-Chain and Off-Chain Mechanisms of DAO-to-DAO Voting" (2026) — arXiv:2603.00708 (preprint) — 61 DAOs and 72 meta-governance relationships. Shifting nodal points and blurring voter composition.
• "Demystifying the DAO Governance Process" (2024) — arXiv:2403.11758 (preprint) — 16,427 DAOs and 122,307 proposals. Over 60% of proposals failed to provide consistent explanations alongside code.
• "Concave is the New Linear: The Impossibility of Anti-Plutocratic DAO Governance" (2026) — arXiv:2605.18990 (preprint) — Impossibility theorem for all balance-based voting rules. Compression functions cannot defeat Sybils.
• "Evaluating DAO Sustainability and Longevity Through On-Chain Governance Metrics" (2025) — arXiv:2504.11341 (preprint) — A repeating pattern of low participation rates and high proposer concentration.
• "Governance Participation in Token-Weighted Voting: Evidence from Decentralized Autonomous Organizations" (2026) — DOI: 10.2139/ssrn.6588294 (preprint) — 1,959 proposals. The length of the voting window has no meaningful relationship with participation; institutional characteristics of each DAO explain most of the variance.
• "'Shadow-Centralization' and Governance Effectiveness in Decentralized Autonomous Organizations" (2026) — DOI: 10.2139/ssrn.6162109 (preprint) — Separating the concentration of holdings from the concentration of exercise. Effects are conditional.
• "The Effect of AI Writing on Governance: Evidence from a $3.5 Billion DAO" (2026) — DOI: 10.2139/ssrn.7284923 (preprint) — The signaling value of the forum collapses discontinuously at a threshold and does not recover.
• "Programmable governance and institutional design in decentralized autonomous organizations: a comparative study" (2026) — DOI: 10.3389/fbloc.2026.1890169 — Three large DAOs independently converge on the same three-tier structure (legal foundation, security committee, tokenholder governance).
• "Democracy for DAOs: An Empirical Study of Decentralized Governance and Dynamic" (2025) — arXiv:2507.20234 (preprint) — Over 3,000 proposals across 20 months. A counterexample showing that participation was maintained or even increased.
• "DAO as Anthropological Experiment: Governance Failure as Subject-Model Conflict" (2026) — DOI: 10.2139/ssrn.6797979 (preprint) — Failures stem from a systematic mismatch between the encoded participant model and real humans. Formal decentralization produces functional recentralization.