Voting as a Deliberative Act — Epoche B2
What Deliberation Takes Away from Condorcet's Theorem The picture to be dismantled A comfortable picture of democratic decision-making has two epistemic engines running in series. First citizens deliberate, and deliberation improves their judgement: arguments are tested, information circulates, and each person ends up more likely to be right than they were. Then citizens vote, and aggregation improves the result again, by the mechanism Condorcet identified in 1785 — errors that are not systematic cancel [1] , and the majority is more reliable than any individual. On this picture the two stages compound, and a democracy that deliberates well and then votes gets the benefit twice. This essay argues that the picture is incoherent, because the two engines run on the same fuel. Condorcet's result requires that votes be independent, and independence is the property that successful deliberation removes. The better the deliberation, the less the count can tell us. Deliberation and aggregation are not complementary stages but rival claimants to a single quantity of epistemic credit, and a theory that assigns it to both has counted it twice. What follows is not that voting is worthless — it is that its value is not epistemic, and the account which gets the division right is Carlos Nino's [2] , in which discussion does the epistemic work and the vote is a device for stopping. What the theorem says, and what it assumes Condorcet's Essai sur l'application de l'analyse à la probabilité des décisions rendues à la pluralité des voix addresses a jury facing a question with a right answer and two options. Suppose each juror votes correctly with some probability p, the same for all, and suppose the jurors vote independently. Then if p exceeds one half, the probability that a majority is correct rises with the size of the jury and approaches certainty as the jury grows. If p falls below one half, the same machinery drives the majority's reliability towards zero. Christian List and Robert Goodin have since shown that the structure generalises [3] : with more than two options, plurality rule tracks the truth under conditions analogous to Condorcet's. The generalisation matters, because it removes the easy objection that real political questions are not binary. What it does not remove is the assumption doing the real work. Independence is not a technical convenience. It is the whole engine. The theorem gets its power from the fact that n independent draws contain n independent pieces of information, so that idiosyncratic errors are unlikely to line up. Remove independence and the arithmetic goes with it. At the limit, if every vote is a copy of every other, the majority is exactly as reliable as one person, however many people there are. What deliberation does to the assumption Now ask what a good deliberative process is for. It is for putting the relevant considerations before everyone, exposing weak arguments, correcting misinformation, and letting those who know something tell those who do not. If it succeeds, the participants finish with a common stock of evidence and a common set of arguments, most of which none of them had at the start. That is a description of manufacturing dependence. Two citizens who have sat through the same discussion and been persuaded by the same argument do not vote independently; they vote as two effects of one cause. Their agreement carries no more information than one of them agreeing with themselves twice, and the count of them registers only how many people the argument reached. The tension is sharpest where the deliberation is best. A process that fails to change anyone's mind leaves independence intact and yields a count worth having, on Condorcet's terms; a process that persuades everyone leaves a unanimous verdict that certifies nothing beyond the persuasiveness of what was said. So the arrangement in the comfortable picture — deliberate first, then aggregate — has the epistemic value of the second stage falling as the first stage does its job. The stages are not in series. They are in competition. The strongest objection The objection to meet is Krishna Ladha's [4] , and it is a result rather than an intuition. In work published in 1992 on the jury theorem, free speech and correlated votes, Ladha showed that the theorem does not collapse the moment independence fails. Provided the average pairwise correlation among votes stays below a bound that depends on individual competence, majority reliability still rises with the number of voters. Some correlation is affordable. The objection then writes itself. Deliberation raises p — participants really do become better judges — while raising correlation by some amount. If the gain in competence outstrips the loss from correlation, the compound picture is vindicated after all, and the choice is not between two engines but between settings on one. Since a well-run discussion plainly does improve individual judgement, the burden lies on anyone claiming the trade is unfavourable. Reply Ladha's result is correct and the objection misapplies it, for a reason it has not conceded: correlation and common cause are different structures, and only the first is what his bound governs. A correlation coefficient is a summary of how often votes coincide. It is agnostic about why. Ladha's condition is a condition on that summary, and it is the right condition for the situation he models — voters who share some background, read some of the same newspapers, and are therefore somewhat more alike than chance would make them. In that situation the votes remain partly independent draws, and the theorem's engine keeps some of its power. Deliberation does not produce that. It produces a shared signal: a specific body of evidence and argument, presented once, on which every participant then conditions. Given the signal, the votes are very nearly determined; what varies between citizens is only their residual disposition to resist a good argument. So th