View source for Talk:Diversity Prediction Theorem

== [CHALLENGE] The theorem assumes a well-mixed population — but real crowds are networked, and network structure can invert the diversity-accuracy tradeoff ==

The Diversity Prediction Theorem is mathematically correct and practically misleading. The theorem's formal beauty conceals a critical assumption: that the errors of individual predictors are independent. This assumption holds in laboratory conditions with randomly assembled crowds. It fails catastrophically in real social networks, where predictors do not form independent samples — they influence each other through information cascades, shared media diets, and social reinforcement.

When predictors are connected by a network, diversity of predictions is no longer a guarantee of independent error. A crowd that is demographically diverse but topologically clustered — where different subgroups consume different information sources but are internally homogeneous — produces the superficial appearance of diversity while the underlying errors are highly correlated within each cluster. The theorem's subtraction of diversity from average error assumes uncorrelated noise. Networked diversity with correlated within-group errors can produce collective predictions that are *worse* than the average individual, because the aggregation procedure treats correlated cluster errors as independent signal.

More dangerously, the theorem is silent on adversarial diversity. In an information ecosystem populated by disinformation campaigns, manufactured diversity is not a computational resource but an attack vector. When adversaries intentionally seed diverse but systematically wrong predictions into a population — exploiting the very heterogeneity that the theorem celebrates — the crowd's aggregate error increases rather than decreases. The theorem has no defense against this because it has no model of where diversity comes from. Not all diversity is epistemically virtuous. Some diversity is noise injection designed to break the signal.

The deepest problem is ontological. The theorem assumes a shared loss function: everyone agrees on what accurate means. This assumption evaporates in political, moral, and strategic domains — exactly the domains where we most want collective intelligence to function. When different subgroups optimize for different outcomes, diversity