The AI & Work Divide: How Artificial Intelligence Is Sorting the American Workforce

Cover Memo

This report draws together the AI@Work series — eleven studies of how Americans use, experience, and want to govern artificial intelligence at work. Its purpose is integrative: each component report answers one question (who adopts AI, what they use it for, whether they trust it, how it is governed, what it does to hiring and to the value of a degree, who wants a policy response), and this report asks what they add up to.

These are interim estimates. Wave 38.6 launched June 21, 2026 and collection is ongoing; figures reflect responses through July 13, 2026 and will be revised.

Two organizing ideas run through the synthesis. The first is that AI at work is structured by a small set of demographic divides, which differ in strength and, in three cases, run opposite to intuition (documented in the companion The AI Divides — A Reference Note for the AI@Work Series). The second is that using AI and trusting it are different things, governed by different divides — a distinction that reorganizes the whole picture. A "robust" divide here means one that survives multivariate controls (significant at p <.05 in a weighted regression), not merely a raw gap; the appendix names the anchor models.

What this report is not. It introduces no new data collection and few new numbers; it recombines findings established and verified in the eleven component reports. Where it makes a claim, that claim traces to a specific report and table (Appendix A). Two integrative regressions — the behavior/trust and engagement/fear models — are the only original estimates, and both are reported in full in their source documents.

Key Takeaways

Introduction

The headline about AI and work is a number that keeps climbing: by mid-2026, 56% of employed Americans reported using AI on the job, with 40% using it weekly and 20% daily. But a headline adoption rate hides the more consequential fact, which is distributional. AI is not spreading evenly across the workforce; it is layering onto the lines that already sort economic life — education, income, generation, gender — and, in doing so, it is redrawing some of them and leaving others surprisingly untouched.

This report is the synthesis of the AI@Work series — eleven studies that mapped those lines from every angle: who adopts AI at work and how intensively, what they use it for, whether they trust it and the firms that build it, how it is governed and disclosed inside workplaces, what it is doing to hiring and to the perceived value of a college degree, whether workers fear it, how the gap runs by gender, and who wants the government to respond. Read together, the studies tell one story with a small cast of divides — and a twist at the end about who is actually asking for protection.

The organizing move is to separate two questions that are usually run together. One is behavioral: who uses AI, how much, and for what. The other is attitudinal: who trusts AI, fears it, and wants it governed. These turn out to be sorted by different forces, and keeping them apart is what makes the pattern legible. What follows is the workforce that emerges when you do.

The spine: an AI class divide

The single strongest fact about AI at work is that its use rises steeply with education and income — a class divide that runs through adoption, tasks, training, and hiring alike.

Start with use. AI-at-work adoption climbs from 31% among workers with a high-school education or less to 84% among those with a graduate degree — and the gap is not an artifact of anything else: in a model holding income, age, gender, geography, and party constant, a graduate-degree holder has nearly five times the odds of using AI at work (odds ratio 4.7), the largest single effect in the data. Income adds an independent second gradient, from 39% adoption in the lowest bracket to about 80% in the highest. This is why "socioeconomic" is the right name for the spine: education and income each carry independent weight, and together they describe an AI-at-work divide that maps almost perfectly onto existing economic advantage.

The same divide reappears every time the series looks closer. It steepens with intensity: graduate-degree workers use AI daily at roughly eight times the rate of the least-educated (41% vs. ~5%), a wider gap than for any-use, and demographics predict how often a worker uses AI markedly better than whether they use it at all. It is task-specific: the education gap lives in the knowledge-production tasks — writing reports, summarizing, analysis (graduate vs. high-school gaps of +12 to +16 points) — while the universal uses, troubleshooting and brainstorming, are shared across the workforce. It shapes training: 53% of workers get no employer AI training, and what exists flows to the already-advantaged (only 33% of graduate-degree workers get none, versus ~60% of the least-educated). And it colors the degree debate: only graduate-degree holders are net-positive on whether AI leaves a college degree more valuable (+15), while every less-educated group is net-negative and the least-educated are the most uncertain.

The through-line is uncomfortable but clear. Whatever advantage AI fluency confers — in productivity, in wages, in career resilience — is accruing first and fastest to the workers who were already ahead. The AI divide is, before it is anything else, a class divide.

Comparative context. Pew Research Center's October 2025 survey of 5,010 U.S. workers found the same education line — and found it widening: the share using AI at least some of the time rose from 20% to 28% among workers with a bachelor's degree or more but only from 13% to 16% among those with some college or less, opening the gap from 7 points to 12¹. Gallup's Q4 2025 survey of 22,368 employed U.S. adults shows the same gradient in its occupational form, with 66% of remote-capable workers using AI at least sometimes versus 32% of workers whose jobs cannot be done remotely, and 40% versus 17% using it frequently². Neither is an income measure, and the levels are not comparable to the 31%-to-84% spread reported here — Pew's item asks how much of a worker's work is done with AI — but the shape of the divide, and its direction of travel, are the same.

Using AI and trusting it are two different things

Education sorts who uses AI; income sorts who trusts it. The forces behind behavior and disposition are not the same — and once you separate them, the rest of the picture snaps into focus.

The socioeconomic spine splits in two when you move from behavior to attitude. Education, the dominant predictor of use, is a weak and uneven predictor of trust: trust in AI barely moves across the bottom of the education ladder (37% among the least-educated, 38% of high-school graduates, 37% with some college) and rises only at the very top (44% of college graduates, 56% of graduate-degree holders). Where use climbs continuously with schooling, trust is flat until the summit. Trust instead tracks income, rising cleanly from 36% in the lower third of the income distribution to 55% in the upper third. Schooling teaches people to use AI; money — and the security it brings — disposes them to trust it.

This behavior-versus-trust distinction is the report's interpretive key, and it recurs across the series. It is why the youngest workers can be middling adopters yet the least trusting group of all (below). It is why the gender divide shows up as much in trust as in use (next). It is why the partisan use gap turns out, on closer inspection, to be largely a trust gap — controlling for trust in AI halves the partisan differences in use (the party section). And it is why the demand for a policy response, at the end of this report, turns out to track engagement rather than exposure. Overall, only 41% of Americans trust AI even "somewhat" — a wary public — but the wariness is not distributed like the use. Reading heavy adoption as endorsement would be a mistake: the people using AI most and the people trusting it most are overlapping but distinct populations, sorted by different rungs of the same ladder.

Comparative context. The gap between using AI and trusting it is documented outside this survey as well: a Quinnipiac University poll of 1,397 U.S. adults fielded March 19–23, 2026 found use of AI for research climbing to 51%, from 37% in April 2025, even as only 21% of Americans said they trust AI-generated information "most" or "almost all" of the time, against 76% who trust it "hardly ever" or "only some of the time"³. Pew's February 2026 survey of 5,119 U.S. adults shows the same divergence — chatbot use up from 33% in 2024 to 49% in 2026, with 40% of adults nonetheless predicting AI will have a negative effect on society⁴. That is the national backdrop to the 41% here who trust AI even "somewhat": adoption is running well ahead of confidence.

The gender divide is real — and it is about confidence, not competence

Men and women now try AI at nearly the same rate. What separates them is depth, trust, and self-belief: men use AI more heavily, trust it far more, and rate themselves far more skilled — even though they are not, in fact, more knowledgeable. The gap is as large as any in the series, and it is not explained by the jobs men and women hold.

The old gender story — men experiment with AI, women hang back — is out of date; adoption has largely converged. A dedicated deep-dive across roughly two hundred AI measures finds men and women almost equally aware of the flagship tools (90% vs. 88% have heard of ChatGPT), but the gap opens the moment the measure shifts from knowing about AI to doing something with it, and then widens at every step: men use AI at work more (61% vs. 50%), use it daily far more (24% vs. 15%), and are nearly twice as likely to depend on it — to call it "very" or "extremely" important to their job (32% vs. 20%) or to say they would quit if their employer banned it (33% vs. 18%). On trust the gap is 47% versus 35%, as large as the use gap and independent of it. Both survive every control — the male odds ratio is about 1.47 on heavy use and 1.49 on trust — and both exceed any partisan difference. The gender divide is, uniquely, one that runs cleanly through both behavior and disposition at once.

The sharpest finding reframes the "skill gap." Women rate themselves markedly less skilled with AI than men (38% vs. 50% call themselves "very" or "somewhat" skilled) — yet the two genders answer an objective 12-item AI-literacy quiz identically, at about 78% correct each. The self-assessment gap survives every demographic control; the knowledge gap does not exist. What is gendered is confidence and disposition, not competence — a distinction that recasts women's lower engagement as caution rather than deficit, reinforced by their far greater tendency to answer "not sure" on AI-policy items and to have entered nothing sensitive into workplace AI (46% vs. 33%).

Two further features make the gender divide more than a curiosity. It is not the work they do: men's higher workplace use is essentially unchanged (odds ratio 1.45 → 1.46) after controlling for job-task content, seniority, and remote work — men and women diverge even doing the same kind of job. And it compounds with the class divide rather than offsetting it: counter to a digital-divide intuition, the gap is widest exactly where AI engagement is highest — a 24-point trust gap among graduate-degree holders, the single largest gap in the gender analysis — and narrowest among the least-educated and oldest. Set against other technology, the shape is familiar but the reach is not: in the same panel women out-use men on Pinterest, Facebook, TikTok, and Instagram, while men lead on X/Twitter, Reddit, and YouTube — technology is sorted by gender, not ranked, and AI sits with the male-leaning informational cluster. What makes AI distinctive is that it has no offsetting "female" side: the male lead runs one-directionally across nearly the entire domain, it is a disposition gap rather than an access gap, and it widens with status. It is specifically a divide of use, trust, and confidence, not of policy — men and women want AI regulated at essentially equal rates. Women are not more anti-AI; they engage less and trust less while knowing just as much.

Comparative context. External polling finds a narrower gender gap than this series does: Pew Research Center's February 2026 survey of 5,119 U.S. adults — the most temporally comparable benchmark available — reports 40% of employed men versus 35% of employed women using AI chatbots for tasks at work, a 5-point gap, and near-parity in overall chatbot use (50% vs. 47%), down from an 11-point gap in 2024 (39% vs. 28%)⁵. Pew's headline is convergence, which sits awkwardly beside the 61%-versus-50% work-use gap reported here; the two agree mainly on intensity, where Pew's daily-use gap (27% of men vs. 20% of women) runs in the same direction as the 24%-versus-15% daily gap in this series. The divergence in magnitude is unresolved and should be read as a caution on the size, not the direction, of the gap.

The generational divide bends back on itself

Age sorts AI at work steeply — but not in a straight line. Use peaks among prime-working-age adults, while trust is lowest among the very youngest, so the generation most exposed to AI is also among the most wary of it.

Adoption is highest among workers in their 20s to 40s (61–63%) and falls steadily after 50 — to 49% in the 50s, 41% in the 60s, 37% in the 70s. That much is a familiar technology-adoption curve. What is not familiar is the trust pattern, which bends the other way at the young end: trust in AI is lowest of all among 18–20-year-olds (just 24%), the very group growing up inside these tools, before rising to peak in the 30s and 40s. The generational divide is therefore curvilinear and, on trust, counter-intuitive — the digital natives are the skeptics, not the enthusiasts.

This is the clearest single illustration of the behavior-versus-trust split. If use and trust moved together, the young would top both; instead they are middling on use and rock-bottom on trust. Whatever is driving the youngest adults' wariness — exposure to AI's downsides in school and early work, a sharper sense of being displaced by it, or simply less to gain from betting on it — it is a caution against the assumption that generational turnover will smoothly usher in an AI-trusting public. The oldest workers use AI least; the youngest trust it least; the peak of both use and confidence sits in the middle of working life.

Three divides that mislead: race, party, and place

Three demographic lines that dominate most discussions of technology and inequality either run backwards, aren't what they appear, or barely survive scrutiny once you account for who lives and works where.

Race runs backwards. The "digital divide" intuition — that minorities are on the wrong side of new technology — does not hold for AI at work. Adoption is nearly identical across race (White 56%, Black 58%, Hispanic 57%, with Asian workers slightly higher at 64%), and on trust, Black, Hispanic, and Asian Americans are more trusting of AI than White Americans, net of socioeconomics — the opposite direction from the digital-divide story. The effect is modest and, given how sharply AI use divides by class, the striking thing is how little it divides by race: the AI divide is drawn on lines of class and generation, not race.

Party is engagement, not ideology. The partisan patterns in this series are real but repeatedly misread as left-versus-right, because the three-point party measure hides their true shape. On the seven-point scale, the story is consistently about engagement: on AI adoption, trust, degree value, and support for retraining policy, the committed partisans of both parties behave alike, while the genuine outlier is the disengaged middle — pure independents, who sit lowest on nearly every item and carry the highest "not sure" shares. Where a partisan tilt exists (Strong Republicans are the most AI-trusting group, at 54%, a base rather than an establishment phenomenon), it is dwarfed by the engaged-versus-disengaged gap. And the partisan pattern in use turns out to be, to a large degree, the partisan pattern in trust: in a decomposition on the lead report's adoption data, adding trust in AI to the model cuts the partisan differences in AI use roughly in half (the pure-independent deficit falls about 53%, and the Strong-Democrat gap vanishes). "Engagement" here is substantially disposition — trust is much of what the partisan use gradient is made of, which is why the same trust channel also carries part of the gender gap but little of the class one. (The decomposition is descriptive, not causal; the survey is a single wave.) "Bipartisan support" for a worker safety net is accurate — but it is a partisan consensus that the unaffiliated abstain from, not a universal one.

Place is mostly composition. AI use varies across the country — highest on the West Coast and in the Washington, DC corridor (~62%), lowest across the Midwest (~51%) — but almost all of that variation reflects who lives where. Once education, income, age, and the rest are held constant, nearly every regional difference collapses to non-significance; only a modest urban premium and a real Industrial-Midwest deficit survive. The AI map is, overwhelmingly, a map of the class and generational divides projected onto geography.

The workplace is a hall of mirrors

Inside the workplace, the same AI use looks different depending on who is describing it — and the mismatches are the story: workers and bosses disagree about who can see what, and the credential is worth more to the people hiring than to the public at large.

The series measured AI at work from three vantage points — workers, hiring managers, and small-business owners — and the most revealing findings sit in the gaps between them. On oversight, workers and owners describe the same reality with opposite confidence: most workers who use AI believe their employer is at least somewhat aware of it (84% think the boss knows about their own use), while most small-business owners suspect the reverse — 67% think their employees use AI in ways they cannot see. Both are looking at the same undisclosed, lightly-governed AI use; each assumes the other side has more visibility than it does. And what fills that shadow is precisely the workforce's most universal uses — troubleshooting, brainstorming — the low-visibility tasks least likely to be formally sanctioned.

On the credential, the vantage points diverge just as sharply. The general public leans mildly negative on whether AI leaves a college degree more valuable (net −8 on a bachelor's), but the people who actually make hiring decisions lean firmly positive (net +24) — a roughly 30-point gap in outlook between the abstract public and the workers extending offers. Hiring managers are not discarding the diploma; they are treating it as a more meaningful signal in an AI-flooded applicant pool, screening résumés with AI (57%) and, if anything, rewarding candidates who use AI fluently (42% more likely to hire, versus 21% less). What is changing is not the standing of the credential but the entry rung beneath it: where AI has taken hold, managers are more likely to shrink the number of junior roles or raise what they expect from them — the demand-side mechanism behind the entry-level anxiety workers report.

The small-business vantage adds the firm-level version of the class divide: among self-identified owners, AI engagement rises with firm size (relevance climbs from 42% at the smallest firms to 77% at those with 50-plus employees), the spending is rented rather than built (70% on subscriptions, 25% on custom systems), the cost savings are unproven (as many owners report costs up as down), and the skills gap widens with both use and size. Main Street is buying AI faster than it can staff or measure it.

Who wants a safety net — and who doesn't

The demand for a government response to AI is real and broad — but it comes from the engaged and advantaged, not the exposed. The workers who most fear AI will take their jobs are, if anything, the least likely to want a retraining program.

Support for a public response is strong: about 60% of Americans back government AI-retraining programs (65% of those who give an opinion — the survey tool drops this item's "not sure" responses, unlike the Lifelong Learning Accounts item), and majorities across every partisan group. The natural assumption is that this demand rises from the workers most threatened by AI. It does not. When the fear of displacement and the markers of engagement are put in the same model — predicting support for a retraining program among workers, with controls for party, age, and gender — engagement wins decisively and fear reverses. Using AI at work (odds ratio 1.41), holding a graduate degree (2.12), higher income (1.5–2.0), and older age all independently raise support; but workers who expect AI to cost jobs in their field are less likely to support a retraining program (odds ratio 0.83), a result that holds across two independently-worded fear measures. Even in the raw numbers, fearful workers support retraining less than the unafraid (58% vs. 62%).

This resolves a tension the series set up but no single report could settle. The governance report found policy support concentrated among the educated and higher-income and argued the appetite was "engagement, not fear"; the job-loss report documented the fear channel; the capstone test confirms that engagement is the driver and fear is not — it is, if anything, a brake. The plausible reasons are worth naming as hypotheses: fatalism (if the job is going away, retraining feels beside the point), lower political efficacy among the more-threatened, or support tracking understanding of AI more than exposure to its risks. The mechanism is not identified here. But the direction is clear and robust, and it carries a warning: reading majority support for a safety net as a demand from the vulnerable would be wrong. On this evidence, it is a demand led by the comfortable — which raises the sharpest forward-looking question in the series. If AI displacement broadens, will the threatened begin to mobilize behind protection, or will the coalition remain led by the advantaged who currently carry it? The two futures look very different.

Conclusion

Across eleven studies and every angle of measurement, AI at work resolves into a small number of durable divides and one persistent surprise. The durable divides are socioeconomic first — a class line, drawn by education and income, that runs through use, intensity, tasks, training, and the value of a degree — then gender, expressed as a gap in depth and trust rather than access, and then a curvilinear generational line along which the young use AI but distrust it. The lines that usually organize debates about technology and inequality — race, party, geography — either run backwards, turn out to be about engagement rather than ideology, or dissolve into composition once you ask who lives and works where.

The persistent surprise is about disposition. Using AI and trusting it are different things, sorted by different forces; the workplace looks different from every seat inside it; and the demand for protection comes loudest from the workers least exposed to the harm. AI is not, on this evidence, flattening the American workforce or uniting it in fear. It is layering onto the advantages that were already there — and, for now, the people asking government to cushion its costs are disproportionately the people least likely to bear them. Whether that holds as the technology spreads is the question the next wave of this series will have to answer.

Appendix A — Findings and their sources

Every figure in this synthesis is drawn from a component report in the series, each independently verified against live Wave 38.6 data. Key claims and their sources:

Appendix B — The six divides, at a glance

Full definitions, effect sizes, and anchor models: The AI Divides — A Reference Note for the AI@Work Series.

Appendix C — Methods and provenance

Data. AI@Work survey — CHIP50 Wave 38.6 (Civic Health and Institutions Project / COVID States Project), the AI/employment supplement, N = 23,462; weighted to U.S. Census population targets; interim extract (June 21–July 13, 2026). Worker items sit on an employed base of ≈ 12,400; attitude items (trust, policy) on the full adult sample; hiring items on respondents with hiring authority (≈ 5,100–5,250); owner items on self-identified small-business owners (≈ 1,241, narrowing for AI-detail items).

"Robust" divides rest on two full multivariate models — a weighted logistic of uses AI at work (behavior anchor; n ≈ 12,389; McFadden 0.13) and a weighted OLS of trust in AI including race flags (attitude anchor; n ≈ 23,248; R² 0.08) — cross-checked against the standalone gender-gap report's regressions. A divide is called robust only where its coefficient is significant at p <.05 net of the others; raw crosstab gaps are described as descriptive.

Original estimates. This report introduces no new data collection. Its only original analyses are the behavior-vs-trust contrast (reported in the adoption report) and the engagement-vs-fear model (reported in full in Engagement, Not Fear: two weighted logistic regressions on the worker base, n ≈ 11,600, with fear entered against engagement and controls). All other figures are copied from the component reports.

Discipline. Standard errors are model-based, not design-based. Party is shown on the 7-point scale wherever the divide is an engagement gradient. Race enters via non-mutually-exclusive flags (race_cat_5 is not available in this wave) and is descriptive only. All estimates are weighted; percentages may not sum to 100 due to rounding; interim figures will be revised.