Swarms beat central prediction only with diverse information and less-correlated errors

The claim

This applies to reflexive systems of organised complexity, such as markets and institutions. It does not apply to physical systems with fixed laws, such as weather or protein folding, where central prediction can win. In reflexive systems, agent swarms and decision markets answer the organised-complexity objection only under conditions. To pass on local facts, as Hayek's price system does, participants must hold genuinely different local information and be paid for acquiring it. Where their outputs are combined into a forecast, as in a decision market, their errors must also be less correlated than those of the central alternative, measured in effective independent votes. A swarm whose agents run one model on one shared context meets neither condition. It is a central predictor with many hands, and Hayek's argument applies to it again. A third condition binds wherever participants act rather than only forecast: each one's reward must track the quality of the combined outcome, not its own share of a contested resource. Where it does not, as in routing and congestion, decentralised choice loses to central allocation even with complete information (Koutsoupias and Papadimitriou's price of anarchy; Braess's paradox), so no amount of local information rescues it. Decision markets meet this condition by paying for accuracy; swarms that also act on their own payoffs must be checked for it separately.