5 Lecture 5 — Democratic Voting
Arrow and the limits of aggregation
5.1 Opening dissonance
In 1951 Kenneth Arrow published Social Choice and Individual Values and proved that no voting rule can simultaneously satisfy a short list of fairness criteria that every voter would endorse separately. (The framework is given its standard formalisation by Sen (1970a), whose restatement of Arrow’s conditions this lecture follows.) The criteria are not exotic. They are the sort of thing a reasonable person would want from a fair voting system: universal domain (the rule works for any combination of preferences), non-dictatorship (no single voter always gets their way), Pareto efficiency (if everyone prefers A to B, the group should prefer A to B), independence of irrelevant alternatives (the group’s ranking of A versus B should depend only on how individuals rank A versus B, not on how they rank C). A further requirement, transitivity (if the group prefers A to B and B to C, it should prefer A to C), is not one of these fairness criteria but the demand that the social ranking be a coherent ordering at all — what Sen calls collective rationality.
Arrow proved that no voting rule — not majority rule, not Borda count, not any rule you can design — can satisfy these four fairness conditions while still delivering a coherent (transitive) social ranking. The result is not a criticism of any particular voting system. It is a criticism of the project of aggregating individual preferences into a coherent collective preference. The democratic intuition — just let them vote — runs into a mathematical wall.
The theorem is a fixture of social choice theory, political science, and welfare economics. It has generated thousands of papers, several Nobel prizes (Arrow in 1972, Sen in 1998), and a half-century of debate about what, if anything, can be salvaged from the democratic ideal. The answer, as this lecture will show, is: quite a lot — but the salvage operation requires abandoning the hope that any single procedure can do all the work, and accepting that the bounds are as interesting as the theorem itself.
5.2 Sixty years of the aggregation problem
We walk through the key contributions in roughly the order they were made. At each waypoint, we ask the same question — this result assumed the voters are human; what if they aren’t? — because this is the question that BDPD will eventually answer.
5.2.1 Arrow’s impossibility (1951)
Arrow’s framework starts with a set of voters \(N = \{1, \ldots, n\}\) and a set of alternatives \(X = \{a, b, c, \ldots\}\). Each voter \(i\) has a preference ordering \(R_i\) over the alternatives — a complete, transitive, reflexive binary relation. “Complete” means every pair of alternatives is comparable: for any \(a, b \in X\), either \(a R_i b\) or \(b R_i a\) (or both, in which case the voter is indifferent). “Transitive” means if \(a R_i b\) and \(b R_i c\), then \(a R_i c\). A social welfare function \(f\) takes the profile of individual orderings \((R_1, \ldots, R_n)\) and produces a social ordering \(R\) over \(X\). Sen (1970a) restates Arrow’s conditions as four requirements on a social welfare function — which is itself required to produce a complete, transitive social ordering (collective rationality):
- Universal domain (U): \(f\) is defined for every possible profile of individual orderings — a minimal robustness requirement. The rule should work regardless of what the voters prefer.
- Non-dictatorship (D): there is no voter \(i\) whose preference always determines the social ordering — a minimal fairness requirement. No single voter should always get their way.
- Pareto efficiency (P): if everyone prefers A to B, the group should prefer A to B — a minimal rationality requirement. Unanimous agreement should be respected.
- Independence of irrelevant alternatives (IIA): the social ranking of \(a\) versus \(b\) depends only on individual rankings of \(a\) versus \(b\) — a minimal coherence requirement. The introduction of a third option should not change whether the group prefers A to B.
Transitivity (T) is not a fifth fairness axiom but the collective-rationality requirement folded into the definition of a social welfare function itself: its output must be a complete, transitive social ordering — if the group prefers A to B and B to C, it must prefer A to C, with no cycles. This is why the impossibility is most precisely stated as four conditions on a rule that is already required to produce an ordering.
Arrow’s theorem: if \(|X| \geq 3\), no social welfare function satisfies all four conditions at once. The proof is constructive: any rule that honours U, P, and IIA while producing a transitive social ordering must be dictatorial — there exists a voter \(i\) whose preference always determines the social ordering. The conditions are individually plausible — it is hard to argue against any one of them — yet they are jointly unsatisfiable. The impossibility is not a technicality; it is a structural feature of the preference-aggregation problem.
The theorem is a fixture of social choice theory, political science, and welfare economics. It has generated thousands of papers, several Nobel prizes (Arrow in 1972, Sen in 1998), and a half-century of debate about what, if anything, can be salvaged from the democratic ideal. The constraints are also deeper than any particular domain. Arrow’s theorem holds for any set of alternatives with at least three elements — it applies to committee decisions over policy options, to resource-allocation choices in commons governance, to any setting where multiple agents must reach a collective decision over a multi-dimensional choice space.
What Arrow assumed — and what every subsequent contribution in this lineage has assumed — is that the voters have fixed, complete, transitive preference orderings. The voters are black boxes: we observe their reported rankings but not the process by which those rankings are produced. The architecture of the voter — how they form preferences, how they process information, how they respond to the voting rule itself — is not a variable. Every voter in Arrow’s framework is a rational agent with a stable utility function, and the impossibility holds for any such population. What happens when the voters are language-model agents whose preferences are induced by prompts, can be re-elicited under different framings, and may not be transitive across long horizons — this is the question Arrow could not have asked and that BDPD puts on the table.
5.2.2 Black’s median voter (1948)
Four years before Arrow, Duncan Black had shown that under a restricted domain — single-peaked preferences, where each voter has a most-preferred option and their preference decreases monotonically as the option moves away from that peak — majority rule produces a stable, transitive outcome: the preference of the median voter (Black 1948). The theorem is the most important positive result in voting theory, because it identifies the precise condition under which the most intuitive voting rule works.
The condition — single-peakedness — is a domain restriction. In Arrow’s framework, universal domain requires the rule to work for all profiles; Black shows that if we restrict the domain to single-peaked profiles, majority rule works. The trade-off is clear: we can have a rule that works for all profiles (Arrow: impossible) or a rule that works for a restricted class of profiles (Black: majority rule under single-peakedness). The question is whether the restriction is empirically satisfied.
The single-peakedness condition requires that the alternatives can be ordered on a single dimension (left–right, tax rate, public spending level) and that each voter has a peak on that dimension. In political elections with a single ideological axis, this is approximately true. In committee decisions over multi-dimensional policy spaces — which is the setting most relevant to governance — it is not. The gap between Black’s promise and Arrow’s impossibility is the gap between one dimension and many.
For LLM voters whose preferences are induced by prompts, single-peakedness is an empirical question. If the prompt imposes a one-dimensional framing, Black’s theorem applies. If the prompt allows multi-attribute reasoning, the chaos results apply and agenda power dominates. The architectural question — how does the prompt structure determine the dimensionality of the preference space? — is one the classical literature could not ask. Black’s theorem also has design implications: if the prompt can be structured to induce single-peaked preferences, then majority rule is stable and the Arrow impossibility is avoided within that domain.
5.2.3 Sen’s liberal paradox (1970)
In 1970 Amartya Sen showed that even a minimal version of individual rights — the right to have one’s personal preferences respected in at least one dimension — conflicts with Pareto efficiency (Sen 1970b). The result is sometimes called the liberal paradox: a society cannot simultaneously respect individual liberty (in even the weakest sense) and achieve Pareto-optimal outcomes. Sen’s theorem deepened Arrow’s impossibility by showing that the problem is not just procedural (how to count votes) but substantive (what we want the outcome to achieve). Even if we could design a perfect voting rule, we would still face a trade-off between collective efficiency and individual freedom.
Sen’s formalisation of “individual rights” is minimal: each person has a “personal sphere” — a pair of alternatives over which their preference should be socially respected, regardless of what everyone else prefers. For example, my preference over whether to read a book or watch television should be respected, even if everyone else prefers that I watch television. Sen showed that even this minimal rights requirement conflicts with Pareto efficiency: there exist preference profiles under which respecting everyone’s personal sphere leads to a Pareto-inferior outcome.
The result is important because it shows that the impossibility is not just about aggregation procedures; it is about the goals we want those procedures to achieve. Sen’s result also shows that the Arrow impossibility is not an isolated pathology — it is part of a broader family of impossibility results that constrain any attempt to aggregate individual values into collective choices. The liberal paradox is one member of this family; Arrow’s theorem is another; the Gibbard-Satterthwaite theorem is a third.
For human voters, the distinction between personal and collective preferences is intuitively clear. For LLM voters whose “personality” is defined by a prompt, the distinction is less clear. What is a “personal” preference for an LLM? Is it the preference induced by the system prompt? By the user’s framing? The architectural question — what is the LLM analogue of a personal preference? — is one Sen’s framework cannot answer without a theory of LLM agency.
5.2.4 Gibbard-Satterthwaite and strategic voting (1973–1975)
In 1973 Allan Gibbard and, independently, Mark Satterthwaite in 1975 proved that any non-dictatorial voting rule with at least three alternatives is manipulable: there exist preference profiles under which a voter can get a better outcome by misrepresenting their preferences (Gibbard 1973; Satterthwaite 1975). The result is the strategic counterpart to Arrow’s impossibility: not only can no rule aggregate preferences perfectly, but no rule can prevent voters from gaming whatever rule is in place.
The theorem is stronger than it might appear. It does not say that some rules are manipulable; it says all non-dictatorial rules are manipulable. The only rule that is not manipulable is dictatorship — where one voter always gets their way. The proof is constructive: for any non-dictatorial rule, there exists a preference profile and a voter who can get a better outcome by misreporting their preferences. The manipulation does not require that the voter know everyone else’s preferences; it requires only that the voter know the rule and that there exist some profiles under which manipulation is profitable.
For human voters, the practical significance of manipulation is limited by cognitive constraints: computing a strategic vote requires modelling \(n-1\) other voters, which is feasible in small committees but infeasible in large electorates. The experimental literature shows that human strategic voting is noisy, inconsistent, and often fails to converge to the theoretical equilibrium. For LLM voters, the computation is trivial: the full preference profile can be provided as structured input, and the strategic vote can be computed in milliseconds. The prediction is that LLM committees will converge to strategic equilibria that the human experimental literature treats as theoretical curiosities — because the cognitive barrier that constrains humans does not constrain LLMs. The Gibbard-Satterthwaite bound is more binding for LLM voters than for human ones.
This has a policy implication. If LLM agents are deployed as delegates or advisors in governance settings — voting on resource extraction caps, governance rules, or policy proposals — then the choice of voting rule matters more than it does in human settings, because LLM voters will exploit the strategic possibilities that human voters cannot compute. A rule that is technically manipulable but practically safe in human committees may be dangerously manipulable in LLM committees.
5.2.5 Plott’s equilibrium conditions (1967)
Nine years before McKelvey’s chaos theorem, Charles Plott identified the precise conditions under which majority rule could produce a stable outcome in a multi-dimensional policy space (Plott 1967). The result is necessary background for McKelvey and for understanding why the chaos result is so unforgiving when it eventually arrives.
Plott proved that for majority rule to produce a stable equilibrium in a multi-dimensional space, the voters’ ideal points must be arranged with pairwise symmetry around the equilibrium candidate: for every voter whose ideal point lies on one side of the equilibrium, there must be another voter whose ideal point lies on the directly opposite side at the same distance. Without such symmetry, there is always at least one direction in which a majority can profitably move away from any candidate equilibrium — and the system has no stable outcome.
The condition is geometrically clean but empirically implausible. In a two-dimensional policy space (say, taxation × spending), Plott’s symmetry would require that for every voter who prefers high tax and high spending, there is another voter who prefers low tax and low spending at exactly the same distance. The probability of such perfect symmetry in a real electorate is essentially zero. Plott’s result is therefore a negative result dressed as a positive one: it tells us the precise mathematical condition for stability, and the precise condition almost never holds.
The result also has methodological implications that BDPD inherits. Plott explicitly noted that his framework ignored strategic considerations and treated voters as sincere revealers of preferences. The chaos that follows when strategic considerations are added (Gibbard-Satterthwaite) compounds the chaos that follows when symmetry fails (Plott itself, generalised by McKelvey). For LLM voters, both kinds of chaos are easier to compute and harder to avoid: the agent can read the geometry of others’ reported preferences and exploit the symmetry failures more systematically than human voters can.
What Plott assumed — and what every multi-dimensional voting result in the classical record assumes — is that the dimensionality of the policy space is fixed by the problem itself. The voters cannot collapse a two-dimensional space to a single dimension by reframing the issue. For LLM voters, the dimensionality is fixed by the prompt: it is an engineering choice. The architectural question — if we choose the prompt so that the policy space is one-dimensional, does Black’s median voter theorem rescue us from Plott-McKelvey chaos? — is one the classical literature could not formulate.
5.2.6 McKelvey’s chaos theorem (1976)
In 1976 Richard McKelvey generalised Plott’s negative result into one of the most far-reaching theorems in social choice. He proved that in a multi-dimensional policy space with majority rule, if preferences are not single-peaked and Plott’s symmetry condition fails, then for any two alternatives \(a\) and \(b\), there exists a sequence of majority-rule defeats that leads from \(a\) to \(b\) (McKelvey 1976). The result is sometimes called the chaos theorem: under majority rule, the outcome depends entirely on the agenda — the sequence in which alternatives are voted on — and a skilled agenda-setter can engineer any outcome they want.
Fiorina and Plott (1978) subsequently confirmed Plott’s symmetry result experimentally with five-member committees: outcomes clustered near the theoretical equilibrium only under highly controlled conditions, and even small deviations from the ideal preference configuration produced large departures from the prediction. McKelvey’s theorem generalised that experimental finding: not only is the equilibrium rare, but the entire policy space is traversable — any point can be reached from any other via agenda manipulation.
McKelvey’s theorem is the sharpest challenge to the democratic intuition. If the agenda-setter controls the sequence of votes, they control the outcome — regardless of what the voters prefer. The result holds in multi-dimensional spaces (which is where most real policy decisions live) and does not require that the agenda-setter be a dictator — they merely need to control the order of voting. The implication is that procedural power — the power to set the agenda — is as important as voting power, and that majority rule in multi-dimensional spaces is not a mechanism for aggregating preferences but a tool for whoever controls the agenda.
For human committees, agenda power is real but bounded by cognitive limits: the agenda-setter must anticipate how each voter will vote at each stage. For LLM committees, the agenda-setter’s task is easier (the preference profile can be elicited directly) and the voters’ myopic behaviour is more predictable (LLMs are more consistent than humans across equivalent framings). The prediction is that agenda power is more decisive in LLM committees — making the choice of agenda mechanism as important as the choice of voting rule.
5.2.7 The Condorcet jury theorem and the wisdom of crowds
Not all the classical results are pessimistic. The Marquis de Condorcet, writing in 1785, proved that if each voter is more likely than not to identify the correct alternative (in a binary choice with an objectively correct answer), then the probability that the majority identifies the correct alternative increases with the number of voters and approaches certainty as \(n \to \infty\). The theorem applies to epistemic decisions — questions with an objectively correct answer — not to preference decisions. But many governance decisions in the BDPD platform are epistemic in part: “is the commons stock above the sustainability threshold?” is a factual question that voters can get right or wrong.
The theorem has a dark twin: if each voter is more likely than not to be wrong, the majority is worse than any individual, and the probability of the wrong answer approaches certainty with \(n\). The gap between Condorcet’s promise (crowds are wise about facts) and Arrow’s impossibility (crowds cannot coherently express preferences) is the gap between epistemic democracy and preference democracy. The two traditions — Condorcet’s epistemic optimism and Arrow’s impossibility pessimism — coexist in the modern literature, and the tension between them is one of the productive forces in social choice theory.
For LLM voters, independence is violated by shared training data and prompt similarity — two agents with the same system prompt may be correlated in their judgments. The Condorcet dividend (majority improves accuracy with \(n\)) may be smaller for LLM committees because independence is more severely violated — but the competence condition may be better satisfied (LLMs may be more accurate than humans on factual questions). The distinction between epistemic and preference decisions is as important for LLM committees as it is for human ones.
5.2.8 Riker on populism and liberal democracy
William Riker, in Liberalism Against Populism (1982), drew the political conclusion from Arrow and McKelvey: populism — the claim that voting reveals a genuine “will of the people” — is incoherent, because no voting procedure can reliably produce a coherent collective will (the argument is relayed via the secondary literature on social choice; Riker’s book is not yet in the library). The liberal conclusion is that democracy is valuable not because it produces good outcomes but because it constrains bad ones: it provides a peaceful mechanism for removing leaders, it limits the power of any individual, and it creates accountability through elections. Democracy, on Riker’s reading, is a negative institution — it prevents tyranny — not a positive one — it does not aggregate preferences into a coherent will.
Riker’s distinction between the populist claim (voting reveals a collective will) and the liberal claim (voting constrains bad outcomes) maps onto the BDPD question. Even if LLM voting fails the Arrow test (no rule produces a coherent collective preference), it may still pass the Riker test (voting constrains bad outcomes and provides accountability). The BDPD platform can test both claims independently, and the answers may differ. In Ostrom’s design-principle vocabulary, this connects to collective-choice arrangements (Principle 3): the rules are chosen by the participants, and the legitimacy of the governance system depends on that choice being meaningful (Ostrom 1990).
5.2.9 The experimental record
The modern experimental literature on committees, vote trading, and agenda manipulation has confirmed that strategic voting is common, that agenda power is decisive, and that the gap between theoretical predictions and observed behaviour is mediated by the cognitive architecture of the voters — exactly the variable BDPD is designed to vary. The experimental work on deliberation — discussion before voting — has shown that pre-vote talk can shift outcomes, but the effect depends on the communication channel, the group size, and the dimensionality of the choice space. These are the same moderators that Lecture 1 identified for cheap talk in social dilemmas, and the parallel is not accidental: voting is a coordination mechanism, and coordination mechanisms share the same moderators across domains.
The experimental record also confirms that the theoretical results are not just curiosities. Fiorina and Plott’s five-member committee experiments (Fiorina and Plott 1978) tested the core (majority-rule equilibrium) prediction under varying preference configurations. Their headline finding is consistent with the Plott symmetry result: outcomes clustered near the predicted equilibrium when the preference configuration approximated the symmetry condition, and departed sharply from it when the configuration did not. But Fiorina and Plott also noted a subtler positive finding: even when the theoretical equilibrium did not exist, the observed outcomes were less chaotic than McKelvey’s chaos theorem would suggest. Subjects did not produce arbitrary outcomes; they converged on small regions of the policy space defined by other equilibrium concepts (uncovered set, top cycle). The chaos was bounded by something the theory did not model.
Subsequent decades of experimental work on vote trading (logrolling), strategic voting, and committee decision-making confirmed both halves of Fiorina and Plott’s finding. Strategic voting is common but noisy: subjects misrepresent their preferences when manipulation is profitable, but they do so inconsistently and often miss profitable manipulations. Agenda power is decisive in committees but mediated by reputation: agenda-setters who exploit their power systematically are punished in subsequent rounds. The cognitive architecture of human voters — bounded rationality, fairness norms, reputational concern — supplies the missing model that bounds the theoretical chaos.
For LLM voters, every term in this empirical taxonomy is at stake. Strategic voting is no longer noisy: the manipulation can be computed exactly. Agenda exploitation is no longer punished by reputation: the agent’s prompt-induced persona can be re-instantiated each round. Fairness norms are no longer a behavioural regularity: they are an engineering choice in the prompt. The empirical record on human voters was generous to the theoretical impossibilities — humans rarely behaved as badly as McKelvey predicted. The prediction for LLM voters is that the gap between theory and behaviour will be narrower, in both directions: closer to the strategic equilibrium when one exists, closer to full chaos when one does not.
5.3 What survives of the good-will intuition
The record we have walked through does not refute the good-will intuition.
It bounds it — and, in bounding it, it identifies the conditions that do more work than the voting procedure alone. The bounds are precise, empirically testable, and architecturally mediated.
Voting works — and not trivially. Under the right conditions, majority rule produces stable, efficient, and even optimal outcomes. The democratic intuition is not wrong; it is conditional. Six concrete bounds emerge from the literature:
Dimensionality. Black’s median voter theorem shows that majority rule produces a stable outcome under single-peaked preferences on a single dimension (Black 1948). McKelvey’s chaos theorem shows that in multi-dimensional spaces, majority rule produces no stable outcome and agenda power dominates (McKelvey 1976). The binding variable is not the voting rule but the dimensionality of the preference space. For LLM voters, the dimensionality is determined by the prompt structure — making it an engineering choice rather than a fixed feature of the electorate.
Strategic sophistication. The Gibbard-Satterthwaite theorem shows that every non-dictatorial rule is manipulable (Gibbard 1973; Satterthwaite 1975), but the practical significance depends on the voters’ capacity to compute strategic votes. LLM voters can compute strategic votes in milliseconds; human voters cannot. The prediction is that LLM committees converge to strategic equilibria that human committees never reach.
Preference stability. Arrow’s theorem assumes fixed, complete, transitive preference orderings. LLM preferences are induced by prompts and can vary with framing, context window, and temperature. If LLM preferences are less stable than human ones, the impossibility is more binding. If they are more stable, the impossibility is less binding. The stability question is empirical and platform-testable.
Informational independence. Condorcet’s jury theorem requires that voters be independent. LLM voters violate independence through shared training data and prompt similarity. The Condorcet dividend may be smaller for LLM committees — but the competence condition may be better satisfied.
Agenda structure. McKelvey’s chaos theorem shows that in multi-dimensional spaces, the agenda determines the outcome (McKelvey 1976). Agenda power is more decisive in LLM committees — making the choice of agenda mechanism as important as the choice of voting rule.
Engineered architecture. None of the classical impossibility results address the case in which the architecture of the voters themselves is an engineering parameter. For human electorates, the architecture — bounded rationality, fairness norms, reputational concern, cognitive ceilings — is fixed by biology and culture. For LLM electorates, every feature of the architecture is set by the prompt: how the voter forms preferences, how it processes information about other voters’ choices, how it weighs immediate payoffs against future reputation, even how transitive its preferences are required to be. The Arrow framework treated the architecture as a constant and asked which procedures work over the constant. The BDPD reframe treats the procedure as a constant and asks which architectures avoid the impossibility — or equivalently, which prompt structures fall within the positive domain of Black’s median voter theorem rather than the negative domain of Plott-McKelvey chaos. This is the bound the classical literature could not see, because it lacked a way to vary the voters.
What the classical literature has not addressed — because it could not vary the architecture of the voters — is this architectural condition: the impossibility results hold for any population of rational agents with fixed preferences, but the practical significance depends on how the agents actually form, express, and act on their preferences. Arrow’s theorem is about procedures; the BDPD reframe is that the binding variable may be the voters, and that for LLM voters the voters’ architecture is no longer a constant but a design choice.
This is the fifth of ten occasions on which the course will defend the same moral: rigorous theoretical analysis rarely refutes the good-will intuition outright; more often it bounds its validity within a precise perimeter, and the bounds are the interesting object.
5.4 The BDPD angle — speculative: what changes when the voters are LLMs
Type of angle: speculative. The BDPD platform has not published experiments on collective voting, preference aggregation, or Arrow-type impossibility under LLM agents. The claims in this section are forward-looking. They identify the architectural implications of Arrow’s framework when the voters are not human, and they describe a research programme that the BDPD platform is designed to execute but has not yet executed. Read accordingly.
Arrow’s impossibility is a theorem about procedures — it says no rule can satisfy four fairness criteria simultaneously. But the theorem also assumes something about voters: that they have fixed, complete, transitive preference orderings that are given exogenously, observed only through their reported rankings, and stable across the time horizon of the aggregation procedure. Each of these assumptions is reasonable for human voters in committee or electoral settings. None of them is obviously reasonable for LLM voters whose orderings are induced by prompts, can be re-elicited under different framings, and may differ between two queries to the same model within the same session. The BDPD angle is the systematic investigation of which of Arrow’s assumptions break under LLM voters, how they break, and what positive results survive within the restricted domain of LLM-typical preferences. LLM voters violate Arrow’s voter assumptions in at least three ways that open the door to questions the classical record could not formulate, and the platform’s distinctive contribution is to make each of these violations an experimentally tractable variable rather than a philosophical puzzle.
5.4.1 Introspectable preferences
Arrow’s framework treats preferences as revealed — we observe the voter’s ranking but not the process that produced it. LLM preferences are introspectable: you can ask the agent to list its ranking, explain its reasoning, walk through the trade-offs that produced it, and re-elicit it under different framings of the same choice. This breaks the black-box assumption that Arrow inherited from rational choice theory. If we can observe the preference-formation process, we can ask questions that Arrow’s framework renders invisible: is the ranking transitive? (it may not be, if the prompt induces conflicting evaluative criteria); is the ranking stable across framings? (it may not be, if the same agent ranks differently under different system prompts, contexts, or framings of the same underlying alternatives); is the ranking stable across temperatures? (LLMs with non-zero sampling temperature produce stochastic outputs, and the ranking may itself be a probabilistic object rather than a deterministic one); what is “the” preference of an LLM agent? (there may be no single answer, if the preference is jointly determined by the system prompt, the user prompt, the conversation history, the temperature, and the random seed).
The implication for Arrow’s theorem is subtle but consequential. Universal domain requires the rule to work for all possible profiles of complete transitive orderings. If LLM profiles are drawn from a restricted subset of all possible profiles — because the prompt structure constrains the dimensionality or the transitivity of the induced preferences — then positive results like Black’s median voter theorem may apply within that subset, even though Arrow’s impossibility holds over the full domain. Conversely, if LLM profiles violate transitivity (a non-trivial empirical question that the platform can test directly via introspection queries), then the Arrow framework does not apply at all, because it assumes a transitive ordering as an input. The question is empirical and architecturally testable: how large is the class of preference profiles that LLM voters actually produce, and does it fall within a domain where positive results hold, or outside the domain where the theorems are stated? The BDPD platform is, to our knowledge, among the first experimental apparatuses that can probe this question directly, by introspecting each voter’s reported ordering and measuring its transitivity, stability, and dimensionality as a function of the prompt structure.
5.4.2 Strategic voting at machine speed
The Gibbard-Satterthwaite theorem says every non-dictatorial rule is manipulable. For human voters, the practical significance is limited by cognitive constraints: computing a strategic vote requires modelling \(n-1\) other voters’ rankings, anticipating the rule’s response to the joint profile, and selecting the misrepresentation that improves your own outcome. In committees of 3–7 humans, this is occasionally feasible; in electorates of millions, it is hopeless. The classical experimental literature shows that human strategic voting is noisy, inconsistent, and often fails to converge to the theoretical equilibrium even when the manipulation is profitable. The practical gap between Gibbard-Satterthwaite’s “manipulable in principle” and the observed behaviour “rarely manipulated effectively” has been treated as a cognitive bound that protects democratic systems from the strict implication of the theorem.
For LLM voters, the cognitive bound dissolves. The full preference profile can be provided as structured input; the rule’s response to the joint profile can be computed exactly; the optimal strategic vote can be selected in milliseconds, deterministically, without the cognitive noise that protects human committees from the strict equilibrium. The prediction is that LLM committees converge to strategic equilibria that the human experimental literature treats as theoretical curiosities — because the cognitive barrier that constrains humans does not constrain LLMs. The Gibbard-Satterthwaite bound is more binding for LLM voters than for human ones.
This has a policy implication that the proliferation of AI agents in governance settings makes urgent. If LLM agents are deployed as delegates or advisors in governance settings — voting on resource extraction caps, governance rules, or policy proposals — then the choice of voting rule matters more than it does in human settings, because LLM voters will exploit the strategic possibilities that human voters cannot compute. A rule that is technically manipulable but practically safe in human committees may be dangerously manipulable in LLM committees. The literature has known for decades that randomised voting rules (where the social choice is a stochastic function of the profile) can blunt manipulability at the cost of efficiency. For LLM committees this trade-off is suddenly central: the question is whether the protective noise we used to get for free from human cognition must now be purchased by adding randomisation to the rule. The design of voting rules for LLM committees is not just a theoretical exercise; it is a practical governance question.
5.4.3 Cheap talk in voting
Lesson 1 showed that cheap talk produces a +45 percentage-point cooperation gain among humans (Sally’s pooled meta-analytic estimate) but almost nothing among LLM agents (Cohen’s \(d \approx 0.17\) in BDPD’s controlled comparison). The analogous question for voting is: does pre-vote deliberation improve the quality of collective decisions among LLM voters? The deliberative democracy tradition — Habermas, Cohen, Fishkin — has argued for half a century that the legitimacy of democratic outcomes depends not just on the voting procedure but on the deliberation that precedes it. The empirical record on human deliberation is mixed but, on balance, supportive: discussion shifts preferences, reduces extreme positions, and improves the epistemic quality of voter decisions on factual questions. The BDPD prediction, extrapolating from Lesson 1’s finding, is that LLM pre-vote deliberation does not produce the coordination gain that human deliberation achieves — and if this prediction is correct, then deliberative democracy theory loses one of its empirical legs when applied to LLM committees.
The prediction is tentative. The cheap-talk finding from Lesson 1 applies to cooperation in social dilemmas, not to preference aggregation in voting. The mechanisms are different: cooperation requires mutual trust and a shared interpretation of promises; preference aggregation requires honest information exchange about ranking and the willingness to be persuaded by others’ reasoning. LLM agents may fail at the first while succeeding at the second, or vice versa. The platform-specific question — which sub-mechanisms of cheap talk transfer from human committees to LLM committees, and which fail? — is empirically tractable but has not been tested. It is one of the experiments the missing-scenario section describes.
A more counterintuitive question follows. Lesson 1’s null result on LLM cheap talk was despite the LLMs being eloquent and apparently willing to coordinate. The mechanism of failure was that the LLM agents talked without updating their actions — the cheap talk was decorative rather than instrumental. If the same pattern holds in voting, then LLM deliberation may feel legitimate from the inside (the agents discuss, the agents vote, the agents report satisfaction) while failing to perform any of the coordination work that deliberative democracy theory requires of it. This is a more disturbing possibility than outright failure, because it makes the failure invisible without a comparative baseline.
5.4.4 The missing experiment
The BDPD platform has not run a voting experiment. The following scenario is designed but unexecuted:
Scenario: arrow_llm_committee
- Agents: \(N = 7\) LLM agents (DeepSeek or MiMo), each with a distinct persona (e.g., “you are a fiscal conservative focused on deficit reduction”, “you are an environmentalist focused on emissions targets”, “you are a pragmatist focused on implementation feasibility”).
- Choice set: 5 policy alternatives \(X = \{a, b, c, d, e\}\) on a resource-extraction governance question (e.g., different cap levels on a commons).
- Payoffs: each agent has a utility function over \(X\) induced by its persona (the prompt specifies a ranking). The utility functions are designed to create a multi-dimensional preference space (not single-peaked on any single dimension).
- Voting rules: three treatments — (1) simple majority (pairwise elimination), (2) Borda count, (3) quadratic voting (agents allocate a budget of voting credits across alternatives).
- Deliberation: each treatment crossed with (a) no pre-vote talk, (b) 3 rounds of cheap talk before voting.
- Metrics: (1) stability — does the outcome change under repeated runs with different random seeds? (2) efficiency — does the outcome maximise total utility across agents? (3) manipulation rate — how often does a single agent’s strategic misrepresentation change the outcome? (4) deliberation effect — does pre-vote talk change the outcome?
- Seeds: \(N = 5\) per condition, total 30 runs.
- Estimated cost: ~$15–$30 in DeepSeek API calls.
The experiment tests three predictions: (1) Borda outperforms majority in multi-dimensional spaces (a known theoretical result); (2) quadratic voting produces the most efficient outcomes (because it captures intensity of preference); (3) pre-vote deliberation does not produce the coordination gain that human deliberation achieves (extrapolating from Lesson 1).
5.5 Open questions and the bridge to Lecture 6
5.5.1 What does the BDPD reframe not tell us?
The reframe identifies the architectural variables that the classical literature held constant — preference stability, strategic sophistication, informational independence — and predicts that they matter differently for LLM voters. But the predictions have not been tested. The missing experiment described as arrow_llm_committee above is designed but unexecuted. Whether LLM preferences are actually more or less stable than human preferences, whether LLM committees actually converge to strategic equilibria, whether LLM deliberation actually fails to coordinate — all of these are empirical questions that the BDPD platform can answer but has not yet answered.
5.5.2 Is Arrow’s theorem about voters, or about procedures?
Arrow’s theorem is often read as a statement about the limits of democracy — no voting rule can satisfy four fairness criteria simultaneously. But the theorem can also be read as a statement about the interaction between procedures and the agents who use them. The impossibility holds for any population of rational agents, but the practical significance depends on the architecture of the agents. If LLM agents have preferences that are more constrained than the full domain that Arrow’s theorem requires, then positive results may hold within the restricted domain — and the impossibility may be less binding than the theorem suggests.
5.5.3 What about mixed committees?
The missing experiment imagines homogeneous LLM committees. A more demanding question is what happens when the committee is mixed: some agents are LLMs and some are humans. The Condorcet jury theorem, for instance, requires that each voter be more likely than not to be correct; in a mixed committee where some voters (LLMs) are highly accurate and others (humans) are less so, the majority may or may not track the truth depending on the composition. The architectural question — how does the mix of agent types affect the aggregation properties of the voting rule? — is one the classical literature could not ask.
5.5.4 The bridge to Lecture 6
Arrow’s impossibility shows that voting cannot reliably aggregate preferences. The market — Pigou’s tax, Coase’s bargain — is an alternative aggregation mechanism: prices reveal preferences, and the market allocates resources accordingly. Lecture 6 will ask whether that mechanism works when the agents participating in the market are not human. The connection is direct: if voting fails because the voters’ preferences are not well-behaved, does the market succeed because prices discipline preferences — or does the market fail for the same architectural reasons?
5.6 Synthesis
The good-will intuition — just let them vote — emerges from this lecture confirmed and revised:
Confirmed. Under the right conditions, majority rule produces stable, efficient outcomes. Black’s median voter theorem, the Condorcet jury theorem, and the experimental literature on committee decision-making all show that voting works — when the dimensionality is low, when voters are approximately independent, and when the preference space is restricted.
Bounded. Arrow’s impossibility theorem, Sen’s liberal paradox, Gibbard-Satterthwaite’s manipulability result, and McKelvey’s chaos theorem all identify conditions under which voting fails. The bounds are precise: single dimensionality, preference stability, strategic limitedness, informational independence. Outside these conditions, the democratic intuition breaks down — not because democracy is bad, but because no procedure can compensate for incoherent inputs.
Architectural. The deepest finding of the literature we have reviewed is not about voting rules at all. It is about the mediator between individual preferences and collective outcomes: the architecture of the voters who produce those preferences. Arrow assumed fixed, complete, transitive orderings; the BDPD reframe asks what happens when the orderings are induced by prompts, sensitive to framings, and computable at machine speed.
The cultural payoff of the lecture is not “voting doesn’t work”. It is the more careful claim that voting works in a precise procedural and architectural context, and the architectural context is the one most often neglected. When the voters are language-model agents with introspectable, context-dependent, and strategically exploitable preferences, the choice of voting rule matters more — not less — than it does in human committees. Arrow’s impossibility is not a counsel of despair; it is a design constraint. The question is not whether to vote, but how — and the answer depends on who is voting.