Why Both Sides of the Race Debate Can Be Statistically Right
2026-08-24 · AI-generated insight
These two notes appear to flatly contradict each other, yet both make empirically defensible claims — and seeing how that's possible is more illuminating than picking a winner.
Lewontin's famous finding is real: take any single genetic locus, and the variation within human populations dwarfs the variation between them. From this he concludes racial classification has "virtually no genetic or taxonomic significance." Winegard and Carl point to something equally real: researchers can sort skulls by continental ancestry with high accuracy, sometimes using only a couple of variables. Ordinary people classify by appearance because the differences "actually exist."
The resolution lies in a statistical subtlety Lewontin's argument glosses over — what A.W.F. Edwards later called "Lewontin's fallacy." Any single trait or gene is a poor discriminator between populations, exactly as Lewontin showed. But small between-group differences across many traits are correlated with each other, and correlated information aggregates. Stack enough weakly informative variables together and classification becomes highly reliable, even though each variable alone tells you almost nothing. The skull data Winegard and Carl cite is precisely this: two variables already get you to 80%; add more and accuracy climbs further.
So the honest synthesis is uncomfortable for both rhetorical camps. Against Lewontin: "most variation is within groups" does not entail "groups are indistinguishable" — that's a non sequitur, and populations are statistically identifiable. Against those who read classifiability as vindicating folk race concepts: the fact that ancestry can be inferred from correlated traits says nothing about how much meaningful difference exists, and Lewontin's core finding — that any given individual's genome is overwhelmingly shaped by individual rather than group variation — still stands.
The deeper lesson is about how statistical facts get conscripted into moral arguments. Lewontin moved from a true premise (within-group variance dominates) to a policy conclusion (classification is unjustifiable) via a hidden statistical error. His critics risk the mirror move: from "classification works" to conclusions the classification data don't support. Both notes are filed under Science Against Prejudice — but each shows how easily a correct measurement becomes an incorrect inference when the conclusion is wanted in advance.