The Word 'Randomly' Is Carrying the Whole Argument
2026-08-01 · AI-generated insight
Gould and Lewontin are making the same move, and it is worth noticing exactly where the weight rests. Both invoke the finding that the largest part of human variation lies between individuals, not between groups. Gould reaches for it via "randomly chosen genetic differences"; Lewontin, likewise, via variation measured one locus at a time. The rhetorical payoff is the same—"skin deep" is literal, classification is empty.
But the tell is right there in how the Lewontin note is filed: under Correlation Across Many Variables. That tag is quietly pointing at the counterargument to the very passage it annotates. The statistical claim both authors lean on—that any single, randomly chosen variable barely distinguishes groups—does not settle the question of what happens when you combine many weakly-informative variables that are correlated with one another. This is the substance of A.W.F. Edwards' famous rebuttal ("Lewontin's fallacy"): a hundred features that each individually predict group membership at 55% can, jointly, predict it at near certainty, precisely because their small signals point the same way.
So the two notes agree on a fact and disagree, without knowing it, on its meaning. The per-variable measurement and the many-variable correlation are answers to different questions. The first asks: does this trait mark a race? (No.) The second asks: does the pattern across all traits cluster? (Sometimes, yes.)
What's striking is that this doesn't rescue the old taxonomy Gould is demolishing—the clusters are shallow, recent, and continuous, as his "recent origin of races" point insists. It refines it. Lewontin's social conclusion—that classification is destructive and of no human value—stands entirely on its own moral legs and needs no help from the genetics. The danger is fusing the two: letting the statistical argument do the moral work. If someone later shows the correlations cluster more than "randomly chosen" differences suggest, you don't want your ethics to have been staked on the null result. The commonplace book has, in a single filing tag, flagged the load-bearing assumption.