Making it up on Volume: Are Larger Groups Really Smarter?

ABSTRACT

Hélène Landemore's Democratic Reason offers a new justification for democracy and for broad-based citizen participation, appealing to the “emergent” intelligence of large, diverse groups. She argues that ordinary citizens should rule as directly as possible because they will make better informed, more intelligent decisions than, for example, appointed officials, councils of experts, or even elected representatives. The foundation of this conclusion is the premise that “diversity trumps ability” in a wide range of contexts. But the main support for that claim is merely a series of computer experiments that are strongly biased toward that result and tell us essentially nothing about decision making in real-world political settings. Moreover, Landemore's analyses of alternative forms of rule (“rule by the one, the few, and the many”) deal only in abstract comparisons between sharply distinguished ideal types. Among other difficulties, they entirely overlook the central consideration in such comparisons: the relative ability of any decision-making process to go beyond stereotyped, intrinsic strategies and integrate multiple sources and varieties of information. In the end, Landemore's claims for the superior intelligence of broadly participatory forms are thus not supported by credible evidence.

Notes

1. Landemore distinguishes justifications based on values, interests, and knowledge (ch. 1).

2. I prefer to avoid Landemore's term epistemic properties which (despite the currency of the modifier) strikes me as vague, indirect jargon with no obvious advantage.

3. A sophisticated reader would also recognize reasons for doubting that existing practices appreciate those benefits to their full extent. Various psychological biases (in-group preference, confirmation bias, and others) probably do result in excessive conformism in many or most settings.

4. There could be such departments without my being aware of them, although presumably the practice is at least rare.

5. Page's 2007 book also summarizes a number of essentially similar studies that he has done with several collaborators. Landemore does not direct attention to any of the other studies, and I have not examined them.

6. To be clear, the objection is not to all use of computational methods. Although I believe that findings from methods that do not use empirical data should always be identified as such, my main point concerns the design of these particular experiments in relation to Landemore's claims about them. Although Page (2007) provides some of the broad language—including the Diversity Trumps Ability Theorem—he is more cautious in drawing practical implications.

7. Page's formulations of the theorem and the conditions for its application are partly bound up with the technical features of his experiments with Hong and others and do not translate readily to ordinary-language, real-world terms. Even so, Landemore's renderings of the conditions appear to be somewhat loose. Perhaps for that reason, she does not actually attribute the conditions to Page. In a peculiar piece of writing, she presents conditions for Page's theorem with no citation to Page and no indication of authorship, one way or the other.

8. I am supplying some terminology and simplifying some inessential features for brief exposition for a nontechnical audience (and a nontechnical writer). I sometimes focus on one version of the experiments among several to make the description more concrete.

9. For example: If the function happens to have more-frequent-than-expected points where a moderately high value (likely to become a status quo) is followed five integers later by a very high value, then agents whose heuristics include a jump of 5 will perform well for this random draw.

10. Because it is possible that the order of jumps (and thus the individual heuristics) makes a difference with a given random set of values, groups with a complete repertoire of jumps will not have exactly the same average performance.

11. Having two individuals whose heuristics both included one of the better jumps would result in that jump being used more often in cases where the group has more than one jump available. That advantage should be quite marginal.

12. As the number of integers becomes larger, the expected outcomes for different heuristics would approach strict equality.