Who Uses AI at Work: Adoption Across the American Workforce

Cover Memo

This report draws on the AI@Work survey (CHIP50 Wave 38.6), a survey of 23,462 U.S. adults (weighted n = 23,485) fielded June 21–July 13, 2026 to boost statistical power on artificial-intelligence questions. All percentages are weighted to U.S. Census population targets and are population-representative; unweighted respondent counts are reported as a reliability guide only.

These are interim estimates. Wave 38.6 launched June 21, 2026 and data collection is ongoing; the figures in this report reflect responses received through July 13, 2026 and will be revised in a later release.

Self-report caveat. Every figure here is a self-report: what workers say about their own AI use, how often they use it, and how much it matters to their job — not audited activity logs or observed behaviour. Self-reported technology use generally runs above measured use, so the adoption figures are best read as upper bounds, and the comparison to external benchmarks below is framed accordingly.

The report concerns employed adults. The central measure — how often a worker uses AI for their job (ai_work) — was asked of employed respondents, an analytic base of roughly 12,400. Two workplace-context items (how widely AI is used in the workplace; how important AI is to the respondent's job) were asked of the same employed base or, for job-importance, of AI-using workers only; each figure states its base. One item used here, job importance (ai_importance), is stored in reverse of its display order (5 = extremely important … 1 = not important, with a separate 0 = "I do not use AI at work"), verified against the questionnaire and the data.

This is a basic adoption report: it covers the prevalence, frequency, and demographic distribution of AI use at work. Companion AI@Work reports cover workplace governance ("shadow AI") and the perceived productivity effects of AI at work.

This is the lead report in the AI@Work series, and it establishes the two distinctions the rest of the series runs on: the demographic divides that structure AI (documented in the companion "AI Divides" reference note), and the difference between using AI and trusting it. A short section below sets up that behavior-versus-trust distinction; the companion reports carry it forward into what workers do with AI, whether they trust the firms building it, and how they want it governed.

Key Takeaways

Introduction

Artificial intelligence has moved into the American workplace with unusual speed, but measuring how far is harder than it sounds, because the answer depends on where the bar is set. Federal Reserve research based on a national survey by Alexander Bick, Adam Blandin, and David Deming found that about 28% of U.S. workers had used generative AI at work as of August 2024, with roughly 22% using it in a given week (St. Louis Fed). Wave 38.6's mid-2026 data show a higher headline number — 56% of workers report ever using AI for their job — but a regular-use figure much closer to the Fed's: 40% use it at least weekly and 20% daily or more. The gap between "ever tried" and "used this week" is itself part of the story.

What the topline conceals is who those workers are. AI use at work is not spread evenly across the labor force; it is concentrated among the more educated, higher-earning, younger, and urban — a pattern of adoption that tracks existing lines of economic advantage. This report documents the prevalence and frequency of AI use at work and maps its distribution across the standard demographic groups, providing the adoption baseline against which questions of governance and productivity can be read.

A majority of workers now use AI on the job

Most American workers use AI for their work at least occasionally, but intensive, everyday use is still a minority.

Among employed adults, 56% report using AI for their job at least once, while 44% never do. Regular use is meaningful but narrower: 40% use AI for work at least weekly, 20% use it daily or more, and 13% use it several times a day. The workplace context is more tentative still: 35% of workers say AI is not used at all where they work, 33% say a few people use it, 21% say many do, and 11% say almost everyone does — so roughly a third of workers are in workplaces where AI use is already widespread. And even among workers who use AI, it is not yet central to most jobs: only 27% call it "very" or "extremely" important to their ability to do their work, while 31% say it is "not important."

Comparative context. The 56% "ever used" figure sits above the ~28% the Federal Reserve measured for August 2024, but the survey's regular-use figures (40% weekly, 20% daily) are consistent with a year of continued growth from that base, and the difference is partly definitional — "ever used AI for work" is a lower bar than "used generative AI at work" (St. Louis Fed). Part of the remaining gap is likely self-report inflation: these are workers' own accounts of their use rather than audited activity, and self-reported technology use generally runs above measured use — so 56% should be read as an upper bound on adoption, and the weekly and daily figures as the more conservative comparison to the Fed's.

Using AI and trusting AI are different divides

The forces that decide who uses AI at work are not the same as those that decide who trusts it — a distinction that runs through this entire report series.

Everything else in this report describes a behavior: who uses AI on the job. But behavior and belief come apart, and holding the two side by side is the cleanest way to read the demographic divides that follow. Alongside its work items, Wave 38.6 asked all adults how much they trust AI; 41% say "somewhat" or "a lot." (Trust here is the general attitude toward AI; work-use is the worker behavior. The point is not the difference in levels but how differently the two are socially patterned.)

The socioeconomic divide splits in two. Education is the single strongest predictor of whether a worker uses AI — but a weak and uneven predictor of whether they trust it. Trust barely moves across the bottom four rungs of the education ladder (37% among the least-educated, 38% of high-school graduates, 37% with some college) and climbs only at the top (44% of college graduates, 56% of graduate-degree holders). Where use rises steeply and continuously with schooling — from 31% to 84% — trust is essentially flat until the very top. Trust instead tracks income, which rises cleanly from 36% in the lower third of the income distribution to 55% in the upper third. Schooling sorts behavior; money sorts disposition.

Gender is the one divide that runs cleanly through both. Men both use AI at work more (61% vs. 50%) and trust it more (47% vs. 35%), and the trust gap is as large as the use gap and survives every control — a divide of intensity and confidence, not access, since men and women now try AI at nearly equal rates.

Age is where behavior and belief diverge most. Work-use peaks among adults in their 20s to 40s and falls after 50; but trust is lowest of all among the very youngest adults — just 24% of 18–20-year-olds — the group most exposed to AI in school and early work. Familiarity here breeds wariness, not endorsement, a caution against reading heavy use as approval.

Two further divides invert the usual intuition. Race runs backwards from the "digital divide" story: AI use at work is near-identical across race (56–58%, Asian workers slightly higher), and on trust, Black, Hispanic, and Asian Americans are more trusting of AI than White Americans, not less. And party is a matter of engagement, not left versus right: on both use and trust, the committed poles run highest (Strong Republicans most of all) and pure independents lowest — the disengaged middle, not either party, is the outlier (see the party section, which shows that trust in AI statistically accounts for about half of that partisan gap in use).

The same split shows up across every group at once when use and trust are laid side by side (Table 0). Reading down the use columns, the socioeconomic and generational gradients are steep; reading down the trust column, they are muted, except at the very top of the education and income ladders. Two groups make the divergence vivid: the youngest workers (age 18–20) use AI at a middling rate (45%) but trust it least of anyone (24%), while high-income and graduate-degree workers are the only groups where both use and trust are high together.

Daily-use cells marked "~5 +" are lower bounds: one sub-cell (unweighted n < 10) is suppressed. Age 80-and-over is omitted from the use columns (all cells suppressed) but trusts AI at 44%. Race uses self-identified flags (not mutually exclusive); all other groups are mutually exclusive.

Three features of Table 0 organize the rest of the report. First, on use, education and income sort workers most sharply, age is curvilinear (a prime-age peak), and race barely sorts at all. Second, on trust, the ordering is different: income and gender lead, the young are the most skeptical, and racial minorities are more trusting than White Americans — a near-mirror of the use pattern. Third, the two columns of "use" (any use and daily use) tell yet another story from each other — the intensity column divides much harder than the access column, a point the intensity section develops. The rest of this report maps the use side in detail; the behavior-versus-trust distinction, and the divides that structure both, organize the companion reports in the series.

Education is the sharpest divide

A worker's education predicts AI use at work more strongly than any other factor measured.

AI-at-work use rises steeply with schooling: from 31% of workers with a high-school education or less, to 41% of high-school graduates, 49% of those with some college, 65% of college graduates, and 84% of workers with a graduate degree. The gap is not an artifact of other characteristics — in a model controlling for income, age, gender, geography, and party, a graduate-degree holder has nearly five times the odds of using AI at work as a high-school graduate, and a college graduate more than twice the odds. Education is the single steepest gradient in the data.

A steep income gradient

AI use at work roughly doubles from the bottom of the income distribution to the top.

Across the survey's ten household-income brackets, use holds around 37–40% through the bottom four (all households under $35,000), then climbs steadily — 47% at $35–50k, 55% at $50–75k, 65% at $75–100k — before jumping to roughly 78–81% in the top three brackets ($100,000 and up; the full bracket-by-bracket figures are in Table 0). The very top bracket ($200k+) dips slightly below the $150–200k group (79% vs. 81%), a difference within sampling noise; the substantive pattern is a flat lower half and a steep upper half. This gradient persists after controlling for education, meaning it reflects more than just schooling — higher-paying jobs are themselves more likely to involve the kinds of tasks AI tools support. Together, education and income describe an AI-at-work divide that maps closely onto broader economic advantage — and, as the interactions section shows, the two reinforce each other rather than merely adding. (The income_cat_10 variable records household income in ten fixed brackets, from less than $10,000 to $200,000 or more.)

Younger and urban workers lead

AI use at work peaks among workers under 40 and in cities, and declines sharply with age.

Use is highest among workers aged 21–40 (61–63%) and falls steadily thereafter — to 49% among workers in their 50s, 41% in their 60s, and 37% in their 70s. (Workers aged 18–20, many of them students, are lower at 45%.) Geography follows a similar tilt: 62% of urban workers use AI for their job versus 54% of suburban and 48% of rural workers, a gap that survives controls. The age and urban gradients are both consistent with where AI-intensive jobs and early-adopter cultures are concentrated.

A gender gap, and a sharper partisan divide

Men use AI at work more than women; and on the full 7-point party scale, adoption peaks among Strong Republicans and collapses among pure independents — structure a three-way split conceals.

Sixty-one percent of employed men use AI for their job versus 50% of women, a roughly ten-point gap that remains significant after accounting for education, income, age, geography, and party (women's odds about a third lower).

The partisan pattern is where the finer scale earns its keep. A simple Democrat/independent/Republican split shows Republicans somewhat ahead (63% vs. 57% of Democrats and 49% of independents), but it hides two things the 7-point measure makes plain (Table 4b). First, adoption climbs steadily with Republican intensity — from 56% of Lean Republicans to 58% of Republicans to 66% of Strong Republicans, the highest-using group in the entire party spectrum. Second, the low point is not independents as a class but pure independents specifically (43%): partisan "leaners" on both sides resemble their parties (Lean Republicans 56%, Lean Democrats 55%), while the unaffiliated middle sits far below everyone. Democrats are comparatively flat across intensity (Strong Democrats 57%, Democrats 56%, Lean Democrats 55%). Both features survive the full demographic model: relative to Strong Republicans, every other group has significantly lower adjusted odds of using AI at work, with pure independents lowest at roughly half the odds (odds ratio 0.52; Table 5). A supplementary model that adds job-task content leaves the partisan gap largely intact (Table 6) — unlike the same tilt in perceived productivity, which washed out once occupation was controlled — so it is not simply a matter of occupational composition. Its mechanism remains unclear, and it should be read descriptively, not as evidence that politics drives adoption.

One candidate for that mechanism is trust. The partisan pattern in use closely tracks the partisan pattern in trust — Strong Republicans trust AI most and pure independents least (see the behavior-versus-trust section) — and the two are statistically linked. Re-fitting the AI-use-frequency model with trust in AI added as a predictor cuts the partisan gaps roughly in half: relative to Strong Republicans, the pure-independent deficit shrinks by about 53% and the Strong-Democrat gap disappears altogether, while trust itself strongly predicts use and lifts the model's fit from an R² of 0.21 to 0.31 (Table 10). About half of the partisan use gradient, in other words, is accounted for by how much different partisans trust AI; the remainder — including much of the pure-independent trough — is not. Because the survey is a single cross-section, the ordering (partisanship → trust → use) cannot be established here, so this is a descriptive decomposition rather than a causal chain: partisanship and AI trust travel together, and together they explain about half of the partisan differences in use.

For comparison, the 3-point measure shows Republican 63.3%, Democrat 56.8%, Independent/Other 48.8% — collapsing both the Strong-Republican peak and the pure-independent trough that the 7-point scale reveals.

Adoption is even across race

AI use at work barely differs by race or ethnicity — the workforce's real AI divides are drawn along class and generation, not race.

White (56.0%), Black (57.8%), and Hispanic (57.1%) workers use AI on the job at nearly indistinguishable rates — a spread of under two points. The one group that stands apart is Asian workers, at 63.8%, several points above the rest. The evenness runs deeper than the headline: daily use is essentially flat across race — White (21%), Black (21%), Hispanic (21%), and Asian (20%) workers all use AI at work every day at about the same rate — so even the Asian group's edge is concentrated in lighter, occasional use rather than intensive use. This near-absence of a racial gap is itself notable given how steeply AI-at-work adoption divides by education and income: the "AI divide" documented in this report is a divide of class and age, not of race.

The intensity divide is steeper than the adoption divide

The groups that lead on adoption lead even more decisively on daily use — and a worker's demographics predict how often they use AI better than whether they use it at all.

The demographic gaps widen sharply when the measure shifts from any use to heavy use. Among graduate-degree workers, 41% use AI at work daily or more; among workers with a high-school education or less, only about 5% do — roughly an eight-fold gap on daily use, against a smaller (though still large) threefold gap on any use (84% vs. 31%). The same steepening appears on party: Strong Republicans use AI daily at more than double the rate of pure independents (31% vs. 13%), a wider relative gap than the two see on any use. Statistically, demographics explain how often workers use AI far better than whether they do: a weighted model of the full 1–7 frequency scale reaches an R² of 0.21, well above the 0.13 pseudo-R² of the any-use model. Adoption is spreading broadly across the workforce, but depth of use remains concentrated in the same educated, higher-earning groups.

The steepening is not uniform — it is sharpest exactly where the socioeconomic and generational lines run, and negligible elsewhere (daily-use column, Table 0). Income divides intensity even harder than it divides access: about 40–44% of the top three income brackets ($100k+) use AI daily, against 8–11% of the bottom four (under $35k). Age peaks in the 30s (25% daily at 31–40) and falls to roughly 11% by the 60s. Community type shows a moderate urban edge (25% urban vs. 15% rural daily). Gender is a real intensity divide even though it is only a mild access divide: men and women try AI at fairly similar rates (61% vs. 50%), but men use it daily 1.6 times as often (24% vs. 15%) — the gap is in how hard they lean on it, not whether they have touched it. Party is moderate (the Strong-Republican-to-pure-independent daily gap noted above). Race is the exception that proves the rule: daily use is essentially flat — 20–21% for White, Black, Hispanic, and Asian workers alike — so the intensity divide, like the access divide, is drawn on class, generation, and gender, not race.

The pattern holds at the level of whole workplaces, not just individuals. Fifty-eight percent of graduate-degree workers say AI is used by many or almost all of their colleagues, versus 36% of college graduates, 24% of those with some college, 20% of high-school graduates, and just 13% of the least-educated — the same educational gradient, reproduced in which workplaces AI has actually saturated.

"+" marks a lower bound: one daily-use sub-cell for the least-educated group (unweighted n < 10) is suppressed, so the true daily-use share is at least ~5%.

Where the divides compound: two interactions

Most of these divides simply add together — but two of them multiply, so that the workers at the intersection pull further ahead (or fall further behind) than a simple sum would predict.

The demographic gaps described so far are, for the most part, additive: being urban adds about the same premium whether a worker is young or old, Republican or Democrat; the pure-independent trough appears at similar depth in every subgroup. But two pairs of divides interact — the combined effect is larger than the parts.

Education times income. The socioeconomic divide is really two gradients, and they reinforce rather than stack. Among college-educated workers, AI-at-work use climbs steeply with income — from roughly 45% in the lower brackets to 86–88% at the top ($100–200k). Among workers with a high-school education or less, the same climb is muted, running from about 35% to only 49% at the top of the usable range. Fit separately within each group, a top-bracket income advantage (versus the bottom bracket) carries an odds ratio near 6–8 among the college-educated but only about 2 among the non-college — income "pays off" in AI use largely once a worker already holds a degree. The two arms of the socioeconomic divide compound: a high-income graduate is nearly universal in AI use, while high income without a degree confers only a modest bump. Demographics also sort AI use far more sharply among the educated (McFadden R² ≈ 0.12 within the college-plus group versus ≈ 0.04 within the high-school-or-less group). Figure 5d shows the two income gradients fanning apart.

Gender times age. The gender gap in AI use is not constant across the life course; it is widest among the youngest workers and narrows with age. Among 18–20-year-olds, 53% of men but only 37% of women use AI at work — a 16-point gap — and on daily use the gap is roughly threefold (16% of young men versus 5% of young women). By the prime working years it has compressed: at 31–40, 67% of men and 58% of women use AI, a 9-point gap. The mechanism is that young men are the earliest, heaviest adopters — already near the top of the age curve in their late teens and early 20s — while women's use rises steeply into their 30s before the gap closes (Figure 5e). In gender-stratified models, the "prime-age premium" (workers in their 20s–40s relative to the 18–20 baseline) is large and significant for women but essentially flat for men: the age curve has a different shape by gender, not merely a different level.

Read together, the compounding is specific. It is the socioeconomic advantages (education × income) and the youngest men that pull away from what an additive picture predicts; party identity, community type, and gender's interaction with education stay close to additive. The interaction analysis is descriptive — it comes from models fit within subgroups, not from a formal interaction term — and is reported as texture on the main divides rather than as a causal claim.

The divide is about the work itself

Much of the demographic gap in AI use is a gap in what jobs actually involve — and task content is a powerful predictor in its own right.

Workers whose jobs involve data processing or analysis are far more likely to use AI (81%, versus 48% of those whose jobs do not), as are those in marketing or content creation (79% vs. 53%), those who make budget or software-purchasing decisions (85% vs. 52%), and those who manage others (71% vs. 47%). Workers who do manual or physical labor are far less likely (43% vs. 67%). Adding these job characteristics lifts the model's explanatory power by half (pseudo-R² from 0.13 to 0.19) and accounts for roughly 40% of the education and income gradients — confirming that much of the "AI divide" is really a divide in the kind of work people do. But not all of it: education, income, age, gender, and party all remain independently significant even among workers doing the same kinds of tasks.

Geographic variation: mostly a demographic map

AI use at work varies across the country, but almost all of that variation reflects where educated, higher-income, urban workers live — not geography itself.

Descriptively, adoption is highest on the Pacific coast and in the Washington, DC federal-and-tech corridor and lowest across the Midwest. Grouping states into finer regions makes the pattern clear (Table 7): the West Coast (63%) and the "DC Orbit" of DC, Maryland, and Virginia (61%) sit at the top; a broad middle of the Northeast, interior West, and the rest of the South clusters near the national average (55–58%); and the Industrial and Plains Midwest sit together at the bottom (~51%). Two refinements stand out. The West's apparent lead is entirely a Pacific-coast phenomenon — the interior West (57%) is unremarkable. And within the South, the only distinctive area is the DC Orbit: the Deep South (54%) and the border South (55%) are statistically identical, both a touch below average.

But this map is, almost entirely, a picture of composition. When education, income, age, gender, urbanicity, and party are held constant, every region's difference from the national average collapses to statistical non-significance — including the West Coast (odds ratio 1.05) and the DC Orbit (1.17, not significant). The lone exception is the Industrial Midwest, which remains significantly below average even after controls (odds ratio 0.84, p =.01): workers across Ohio, Indiana, Illinois, Michigan, and Wisconsin use AI at work less than their demographics alone would predict — a modest but real regional deficit. Adding geography to the adoption model barely changes its fit (a pseudo-R² gain of well under 0.01), confirming that place adds little beyond who lives there.

At the state level (Appendix B, Table 8), the highest rates are in Washington (67%), New York (64%), Maryland and Delaware (63%), and California and Massachusetts (62%); the lowest are in West Virginia and Wisconsin (42–43%), Oklahoma (44%), and Missouri and Mississippi (45%). Estimates for small-population states carry wide margins of error — up to about ±16 points where fewer than 40 workers were sampled — and should be read with caution; the regional groupings below, each with 780 or more respondents, are far more reliable.

Conclusion

By mid-2026, using AI at work has become a majority experience for American workers, though one still concentrated in occasional rather than constant use and in workplaces that are, more often than not, only partly converted. The more consequential finding is distributional: AI use on the job is layered onto existing lines of advantage. The workers most likely to be using these tools are the more educated, the higher-earning, the younger, the urban, and — more modestly — men and Republicans. Education is the sharpest of these divides, with graduate-degree holders nearly five times as likely to use AI at work as workers with a high-school education. And the divides do not merely stack: education and income multiply — income's payoff in AI use is far larger for workers who also hold a degree — and the gender gap is widest among the youngest workers before women catch up. The concentration of AI use at the top of the workforce is therefore sharper than any single gradient implies, and sharper still for intensive, daily use than for use of any kind.

That pattern matters because it shapes who stands to gain from whatever productivity, wage, or career advantages AI use confers, and who may be left further behind. This report establishes the adoption baseline; the questions of how that use is governed, and what workers believe it does to their productivity, are taken up in companion reports.

Appendix A — Methods

Data source. AI@Work survey — CHIP50 Wave 38.6 (Civic Health and Institutions Project / COVID States Project), the AI/employment supplement. Total unweighted n = 23,462; weighted n = 23,485. Fielded June 21–July 13, 2026, with an oversample of students and adults 18–22; it is nonetheless a broad adult sample. Estimates are weighted to U.S. Census population targets. Data collection was ongoing at the time of this extract — figures reflect responses through July 13, 2026 and are preliminary. Geographic groupings use a derived region_custom recode of state; state-level estimates for low-population states carry wide margins of error (Appendix B, Table 8), and geographic differences are descriptive — they do not survive demographic controls except for a modest Industrial-Midwest deficit.

Measure and universe. The primary outcome is ai_work — "How often do you use AI for your work?" (Never; I tried once or twice; About once a month; About once a week; Multiple times a week; Once or twice a day; Several times a day) — asked of employed respondents (unweighted n ≈ 12,400). "Uses AI at work" collapses all categories above "Never." Workplace penetration (ai_workplace) was asked of the employed base; job importance (ai_importance; reverse-coded, 5 = extremely … 1 = not important, 0 = does not use AI at work) of AI-using workers only. The behavior-versus-trust section additionally uses pol_trust_ai ("how much do you trust AI"; 1 = not at all … 4 = a lot; "trust" = somewhat + a lot), which was asked of all adults (unweighted n ≈ 23,300), not only workers — so its figures describe the general-population attitude and are labeled as such. That section compares the demographic structure of trust against that of work-use; it does not treat the two as sharing a base. Trust figures for every demographic in Table 0 (education, income bracket, gender, age, 7-point party, race, community type) were computed directly from pol_trust_ai on the full adult sample; "trust" sums the "somewhat" and "a lot" categories. Race trust uses the same non-mutually-exclusive race_* flags as race use. All measures are self-reported; no workplace records, activity logs, or employer data were used.

Estimation. Cross-tabulations are weighted shares. The primary multivariate model is one weighted logistic regression (survey-weighted GLM, logit; n = 12,389; McFadden pseudo-R² = 0.13) of "uses AI at work" on the six standard demographics, using the 7-point party scale (party7, referenced to Strong Republican), education, gender, income, community type, and age. A companion weighted OLS on the full 1–7 frequency scale (ai_work; same predictors; R² = 0.21) is used to characterize the intensity divide. A separate logistic model adds job-task content (work_context; McFadden 0.19); that occupation model retains party as the 3-point measure and is reported only as a robustness check (Table 6). Two-way interactions were probed with subgroup-stratified logistic models (Table 9) rather than formal interaction terms, which the estimation tool does not fit: education × income was assessed by fitting the model separately within the college-plus and non-college strata and comparing income odds ratios; gender × age by fitting separately within men and women and comparing age odds ratios. These subgroup comparisons are descriptive. Intensity figures ("daily or more") sum the two highest ai_work frequencies (once–twice a day; several times a day); income is analyzed across all ten income_cat_10 brackets. Standard errors are model-based, not design-based. A gap is described as surviving controls only where the coefficient is significant at p <.05; raw crosstab gaps, and all subgroup interaction contrasts, are descriptive. A trust-mediation check (Table 10) re-fits the 1–7 frequency OLS with and without trust in AI (pol_trust_ai, entered as a linear 1–4 term) and compares the party7 coefficients; because the data are a single cross-section, it is reported as a descriptive decomposition of how much of the partisan use gap co-varies with trust, not as a causal mediation test (the partisanship → trust → use ordering is assumed, not established).

Race/ethnicity. The single 5-category race variable (race_cat_5) was not fielded in this supplement, so race is measured from self-identified binary flags (race_white, race_black, race_hisp, race_asian), which are not mutually exclusive. Race enters the report descriptively only (it is not in the regression models). Cells with unweighted n < 10 are flagged and suppressed (workers 80+, n = 21; the daily-use sub-cell for "some high school or less," reported as a lower bound in Table 4c).

Reproducibility. Figures map to Appendix B tables; values are copied verbatim from the underlying tabulations. Party breakdowns use party7; race breakdowns use the race_* flags. Item wording and skip logic sourced from the Qualtrics instrument via the project's Wave 38.6 verified-updates crosswalk.

Appendix B — Data tables

Race/ethnicity groups are from self-identified flags (race_white/race_black/race_hisp/race_asian) and are not mutually exclusive; the 5-category race variable was not fielded in this supplement. Unweighted n: White 7,776; Black 3,326; Hispanic 1,971; Asian 490.

Every group on the 7-point party scale has significantly lower adjusted odds of using AI at work than Strong Republicans, with pure independents lowest (roughly half the odds). Age 21–30 does not differ significantly from 31–40 once the other factors are held constant (p =.06).

In this job-type model the education and income effects shrink by roughly 40% (the graduate-degree odds ratio falls from about 4.7 to about 3.0; top income bands from ~2.8 to ~2.3) but stay significant, while the gender, age, and party effects are essentially unchanged. This supplementary occupation model retains party as the 3-point measure; the 7-point party estimates in Table 5 are the report's primary partisan results.

Table 9. Interaction checks — subgroup-stratified logistic models of "uses AI at work" (weighted odds ratios; * p <.05; ns = not significant). Each model is fit separately within the stated subgroup; all six standard demographics are included as covariates but only the interacting term is shown.

Panel A — Education × income (income brackets, reference = lowest bracket):

Income's effect on AI use is roughly three times as strong among the college-educated — the socioeconomic advantages multiply rather than add. Model fit is also far higher within the college-plus group (McFadden R² 0.12 vs. 0.04).

Panel B — Gender × age (age groups, reference = 18–20):

Relative to the 18–20 baseline, women's use rises steeply into the prime working years while men's is flat — young men are already heavy adopters — so the gender gap in AI use is widest among the youngest and narrows with age. Gender's interaction with education is milder (the graduate-vs-college odds ratio is 2.37 for men, 1.85 for women); party and community type are stable across subgroups (approximately additive).

All six party gaps are significant at p <.05 in Model A; in Model B every gap except Strong Democrat remains significant. Trust in AI is a strong linear predictor of use (t = 42.8). The same decomposition attenuates the gender use gap by about a third (male coefficient +0.44 → +0.29) but leaves the education and income gradients largely intact — trust accounts for much of the partisan and gender patterns, not the class one. Model A n = 12,389; Model B n = 12,303 (a small number of cases lack a trust response). The survey is single-wave, so this is a descriptive decomposition, not a causal mediation test.