The Contempt Atlas
An open, reproducible look at how class, regional, and political contempt actually moves through American text — and a test of how much of it we imagine.
The question
Survey research keeps finding a perception gap: people guess the other side holds them in far more contempt than the other side reports feeling. We wanted to test that against language. If mutual contempt is overestimated, then a lot of what looks like contempt in text shouldn't be people aiming an insult at all — it should be people reporting that others look down on them, or reclaiming the word with pride.
The metric — AAI
For every contempt term we split each real use three ways: aimed at a target (DIRECT), attributed to others (“they call us flyover country”), or claimed with pride (“damn right I'm a redneck”). That distribution — the Attribution Asymmetry Index — is the connective tissue: it powers the personal test, the term pages, and the network map.
What the data shows so far
The honest answer turned out to be more interesting than the one we set out to prove.
- It depends on direction. When contempt actually crosses between groups it's overwhelmingly aimed — people aren't imagining that. The attributed and reclaimed talk is something groups mostly do among themselves.
- Same word, different room. “Redneck” is thrown as an insult in political subreddits but reclaimed with pride in regional ones. The community changes the meaning more than the word does.
- The tell is in the word. Terms coined as accusations (“bitter clinger,” from a famous quote) are used attributively; ordinary epithets (“yuppie”) are aimed.
The long version, with live numbers, is in the flagship essay.
How it's made
A Python pipeline samples real Reddit comments containing each term, drops author handles at ingest, takes a sentence-sized window of context, and classifies its stance, target, and speaker with an LLM. The current build rests on 6,716 classified observations across 23 terms and 6 groups. Everything is precomputed to versioned JSON and served as a static site; the lexicon, rubric, and aggregates are all published. See the methodology for the full recipe and the source.
Status & honesty
This is dataset version 0.3.0-wave2, and it is deliberately labelled suggestive, not findings everywhere. The stance classifier isn't yet validated against a human gold set; term-matching can only see contempt that contains the word, so it likely undercounts attribution; and several terms are thinly sampled. We show the numbers straight anyway — including where they contradict what we hoped to find. The corpus measures contempt expressed in text, never contempt felt.