The AI Divides — A Reference Note for the AI@Work Series
Purpose. A shared definition of the demographic divides that structure AI use and attitudes, so every report in the series names, ranks, and frames them consistently. "Divide" here means a difference that survives multivariate controls (significant at p <.05 in a weighted regression), not a raw crosstab gap. Three divides are robust and independent; three are real but qualified. Lead with the robust three; treat the qualified three with the noted care.
The six divides at a glance
| Divide | Verdict | Evidence, net of controls | How to treat it in reports |
|---|---|---|---|
| Socioeconomic (education × income) | Robust — primary | Education leads behavior: grad-degree OR 4.7 on AI use at work. Income leads attitudes: top brackets the strongest positive coefficient on trust (+0.30). | The series backbone. Name it "the AI socioeconomic divide" — the combined term is accurate because education drives who uses AI and income drives who trusts it. Every report should locate its finding on it. |
| Gender | Robust — co-primary | Male OR 1.47 on heavy use, 1.49 on trust (both p<.001); trust coefficient +0.165 even net of race. Null on wanting regulation (OR 1.04, ns). | Co-lead with socioeconomics. Frame as an intensity-and-trust divide, not access (men and women now try AI at near-equal rates). Don't manufacture a gender gap on policy/regulation direction — there isn't one. |
| Generation (age) | Robust — third | AI use falls steeply after 50 (OR 0.56 → 0.27 across 51–60 to 71–80). Trust is lowest among 18–20 (−0.40), peaks in the 30s–40s. | Handle as curvilinear, not linear. "Young = AI-native" is wrong on attitudes: the youngest adults are the least trusting. Prime-working-age is the adoption peak. |
| Party | Robust but structured | On attitudes, large: pure independents and Lean Democrats lowest on trust (−0.43), Strong Republicans highest. On use, modest (Strong Rep peak 66%, pure independents 43%). | Real, but not a left–right gradient. Use party7, not party3 — the story is the pure-independent disengagement trough and a strong-partisan pattern the 3-point scale hides. Keep party3 only for genuinely linear items (e.g., job-loss fear). |
| Race (indicator flags) | Robust but reverse & modest | Net of SES, Black +0.18, Hispanic +0.07, Asian +0.07, White −0.07 on trust. On any-use, near-even (White 56 / Black 58 / Hisp 57 / Asian 64%). | The opposite of "digital divide" intuition: minorities are more positive/trusting/higher-intensity, not less. Modest in size and easy to misread — state the direction explicitly. race_cat_5 is unexposed in the AI@Work survey; use the race_* flags (non-mutually-exclusive) and keep race descriptive. |
| Geography (urban/rural) | Weak — mostly composition | An urban premium survives (use OR 1.42, trust +0.12); the rural–suburban gap largely washes out under controls. | The weakest robust signal. Report state/region maps as descriptive — most geographic variation is who lives where, not place itself. Don't over-weight. |
How to use this
- Foreground the robust three — socioeconomic, gender, generation — as the series' organizing axes. A report that leads with party, race, or geography as its main divide is usually leading with a weaker or composition-driven signal.
- Distinguish behavior from attitude. The socioeconomic divide splits: education predicts use, income predicts trust. Say which outcome you're describing before naming the driver.
- Descriptive vs. independent. A raw crosstab gap is descriptive; call a divide "meaningful," "independent," or "survives controls" only when a regression coefficient backs it (p <.05). This is the project's significance discipline applied to demographics.
- Party needs the 7-point scale. Default to
party7(or at least split pure independents) for policy/attitude outcomes;party3only for simple left–right gradients. - Race is a reverse divide. Frame minority AI positivity as the finding it is, not as a shortfall; never imply a White advantage the data don't show.
What "robust" rests on
Verdicts are anchored in two full multivariate models on the AI@Work survey AI & Employment supplement (weighted; all six demographics plus race flags), cross-checked against the standalone gender-gap report (Waves 35/38):
- Behavior anchor — weighted logistic regression of uses AI at work on the demographic set (n ≈ 12,389; McFadden 0.13).
- Attitude anchor — weighted OLS of trust in AI (
pol_trust_ai, 1–4) on the demographic set plus race flags (n ≈ 23,248; R² 0.08). - Gender anchor — the AI gender-gap report's weighted logistics: regular use OR 1.47 (W35, n 9,370), trust OR 1.49 (W38, n 31,683), regulation OR 1.04 ns.
Standard errors are model-based, not design-based. Race enters via non-mutually-exclusive flags (race_white/black/hisp/asian) because race_cat_5 is not available in the AI@Work survey. All figures are weighted estimates; interaction terms (e.g., gender × education, where the gender gap is widest) are not fit by the current tooling and are reported descriptively where used.
Prepared 2026-07-21 · AI@Work series.
AI@Work survey — CHIP50 Wave 38.6 (Civic Health and Institutions Project / COVID States Project). All figures are weighted estimates.