Judging by the Nose: When Refusing to Aggregate Evidence Becomes a Method
2026-08-31 · AI-generated insight
There's a structural rhyme running through these notes that goes deeper than their shared topic. Lewontin's famous argument — that most genetic variation lies within populations rather than between them — turns out, per Winegard and Carl, to depend on examining one locus at a time. Their analogy is memorable: you can't sex a face by nose size alone, but consider all features together and the classification becomes trivial. The statistical error is refusing to aggregate: keeping each piece of evidence isolated so that no pattern can emerge.
Now look at the Coon anecdote. The AAA members who voted to denounce Putnam's book hadn't read it; most hadn't heard of it. They too were judging without aggregating — condemning a work on the basis of a single signal (its reputation among a few upset colleagues) rather than the full evidence available to anyone willing to look. In both cases, the epistemically thin procedure wasn't an accident. It produced the conclusion that felt morally necessary, and a more thorough procedure risked producing one that didn't.
This reframes what's happening in Lewontin's own passage. Notice how quickly it moves from a statistical finding to "no justification can be offered" for racial classification — the empirical claim was load-bearing for a moral one. That's precisely the arrangement Coon's story warns against: once a scientific claim is doing moral work, the incentive to scrutinize its methodology inverts. Correcting Lewontin's statistics stops being a technical improvement and starts feeling like an ethical transgression, which is presumably why the correlation-across-loci point took decades to become common knowledge despite being, as the face analogy shows, intuitively graspable.
The deeper lesson isn't about race at all. It's that "condemnation without examination" and "analysis without aggregation" are the same failure mode at different scales — both are ways of arranging your inputs so the inquiry can't surprise you. A field that tolerates one will tend to produce the other.