The AI Gender Gap: How Men and Women Diverge Across Every Corner of Artificial Intelligence

Cover Memo

This report is a deep dive into gender differences in artificial intelligence, drawing 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. It examines gender differences across the full battery of AI items in the supplement — awareness, tool-by-tool use, workplace use and intensity, self-rated and objective skill, purposes, affect, trust, policy attitudes, workplace consequences, disclosure, training, and hiring — roughly two hundred variables in all.

Three analytic passes sit behind it. First, a descriptive sweep of every AI variable, reporting the weighted men-vs-women difference on the correct base for each item (all adults; the ~12,400 employed; ~17,500 AI users; or smaller filtered blocks, each labeled). Second, an intersection analysis measuring how the gender gap varies across age, income, education, race, urbanicity, and 7-point partisanship. Third, a mediation test asking whether men's greater workplace AI use is explained by the kind of work men and women do — using survey-weighted regressions with and without controls for job-task content, remote status, and seniority.

Gender is the vendor two-category measure (Male/Female). All percentages are weighted; unweighted counts are reported only as reliability guides. Where a finding rests on a regression it is labeled as surviving controls (p <.05); otherwise gaps are descriptive. This is a companion analysis in the AI@Work series, extending the behavior-versus-trust framework of the lead report into a single demographic axis.

Key Takeaways

Introduction

Artificial intelligence is diffusing through American life along the familiar contours of advantage — education, income, age. This report isolates a different axis: gender. Using the AI@Work survey's unusually complete AI battery, it asks not whether a gender gap exists on one or two headline questions, but how men and women differ across the entire surface of AI — the tools they have heard of, the tools they use, how intensively, for what, how they feel about it, whether they trust it, whether they want it governed, and what it is doing to their work.

The answer is unusually consistent, and that consistency is itself the finding. On almost every one of roughly two hundred measures, men report more engagement than women. The interesting questions are therefore not whether men lead but where the gap is widest, what it is made of, and what explains it — in particular, whether it reflects the different jobs men and women hold, or something that persists even when the work is held constant. Two results reframe the usual "women are behind on AI" story: the skill gap is a gap in confidence rather than knowledge, and the gap is largest not among the marginalized but among the most advantaged.

Is AI special? How the gap compares with other technology

The male lead on AI is not a story of women using less technology — in the same panel, women out-use men on several major platforms. Technology use is sorted by gender, not ranked; AI simply sits on the male side of a split that long predates it.

Set the AI findings against social-media platform use measured in the same survey panel and the "women are behind on tech" reading falls apart. Women lead — often by a wide margin — on the social and visual platforms: Pinterest (40% vs. 20% for men), Facebook (82% vs. 76%), TikTok (50% vs. 43%), and Instagram (57% vs. 54%). Men lead on the informational, public-argument, and technical platforms: X/Twitter (34% vs. 19%), YouTube (78% vs. 72%), Reddit (28% vs. 23%), and LinkedIn (23% vs. 21%). Pew Research Center documents the same split independently — women far more likely to use Pinterest (50% vs. 19%), Instagram, and TikTok. On this map, AI belongs squarely to the male-leaning cluster: its +10-to-+12-point gaps in workplace use and trust resemble X/Twitter and Reddit, not Pinterest.

So the direction of the AI gap is unremarkable — men have long led on the technical and discourse side of consumer technology. Three features are what make AI distinctive:

• It does not cancel out. Overall social-media use is roughly even between the genders precisely because the female-leaning and male-leaning platforms offset one another. AI has no offsetting "female" side — men lead on nearly every one of the ~200 measures in this report. A one-directional gap across an entire technology domain is unusual.
• It is a disposition gap, not an access gap. A classic technology divide is about reach; the AI gap is not. Women match men on ChatGPT awareness and on objective AI literacy. What is gendered is confidence, enthusiasm, and trust — the same character as the male-leaning "opinion" platforms, where the barrier was never capability but propensity to engage.
• It widens with status. Platform-gender splits are fairly flat across the class ladder. The AI gap is largest among the most educated and affluent — the feature with the sharpest inequality implications, and the one most particular to AI.

A gap that spans the entire AI lifecycle

Men and women start from near-parity on basic awareness, then diverge at every subsequent step — using AI, using it intensively, trusting it, and depending on it.

Awareness of the flagship tools is essentially equal — men and women are within a few points of each other on having heard of ChatGPT and Google Gemini. (These awareness figures come from the survey's ai_know battery, which returned invalid values when broken down by gender in analysis — see Methods — so they are reported with caution; the near-parity is independently corroborated by Pew, which finds identical ChatGPT use across genders, 44% each.) But the moment the measure shifts from knowing about AI to doing something with it, a gap opens and then widens. Men are more likely to use AI for their job at all (60.7% vs. 50.5%), to use it daily or more (24.0% vs. 15.3%), to have used an autonomous "AI agent" (46.8% vs. 37.8%), and to say AI's impact on their own job is "only positive" (19.9% vs. 13.1%). It is widest on measures of dependence: men are nearly twice as likely as women to say they would be likely to quit if their employer banned AI (32.7% vs. 17.5%) and to call AI "very" or "extremely" important to their job (32.1% vs. 19.6%). The gap is not a wall between users and non-users; it is a gradient that steepens with depth of engagement.

Comparative Context

The direction and much of the texture match Pew Research Center's June 2026 report *The gender gap in AI*, based on a survey of 5,119 U.S. adults (fielded February 17–23, 2026). Pew likewise finds men and women nearly converged on overall chatbot use (50% vs. 47%) but men still ahead on daily use (27% vs. 20%), on confidence using chatbots (22% vs. 15% "extremely/very confident"), on perceived productivity gains (35% vs. 25%), and on expecting AI to help them personally (29% vs. 17%), while women are more likely to say AI is moving too fast (68% vs. 58%). Pew records identical ChatGPT use across genders (44% each) — echoing the near-parity in awareness measured here. What the AI@Work survey adds is breadth: the same gap runs not across a handful of items but across the entire instrument — into workplace intensity, dependence, application-by-application support, and the gap between perceived and objective skill.

Confidence without competence

Women rate themselves far less skilled with AI than men do — but answer an objective AI-literacy quiz exactly as well. The skill gap is in confidence, not knowledge.

Asked to rate their own skill with AI tools, 50.1% of men but only 38.1% of women call themselves "very" or "somewhat" skilled — a 12-point gap that survives controls for age, education, income, party, and community type (men score about 0.2 points higher on the 1–4 proficiency scale net of all of them, p <.001). Yet on an objective, 12-item true/false AI-literacy quiz, men and women score essentially identically: about 78% correct each (77.8% vs. 77.7%). On eleven of the twelve items the gap is under two points; the one exception runs toward women (they are 6 points more likely to correctly reject "AI always gives factual answers").

The divergence is clean and important: women's lower self-rated skill is not tracking lower actual knowledge. It is a confidence gap. (One caveat cuts the other way: on a separate six-item quiz about AI data centers — physical-infrastructure facts — men do score genuinely higher, by 9–14 points per item, largely because women answer "not sure" far more often. The confidence-without-competence pattern is specific to general AI literacy, not to every knowledge domain.)

An enthusiasm gap, not a fear gap

Men feel markedly more positive about AI; women's higher negativity is real but narrower, concentrated in acute distress rather than general anxiety.

Respondents rated how well eleven emotions describe their feelings about the rise of AI, on a 1–10 scale. Men out-score women on all four positive emotions by close to a full point: excited (5.08 vs. 4.14), optimistic (5.12 vs. 4.21), enthusiastic (4.90 vs. 4.02), and curious (5.78 vs. 5.10). These are the largest gaps in the entire affect battery. Women score higher on the negative emotions, but only two clear the threshold for a notable gap — scared (5.19 vs. 4.81) and overwhelmed (4.85 vs. 4.51). The more cognitive negatives — concerned, anxious, angry — differ by less. Notably, "concerned" is the single highest-rated emotion for both genders (6.13 for women, 5.96 for men): shared wariness is the common ground beneath a large enthusiasm gap.

The framing matters. It would be easy to read the gender gap as women being more afraid of AI. The data say something more specific: women are not dramatically more anxious in the abstract; they are less enthusiastic, and more likely to feel personally scared or overwhelmed by the pace of it.

Trust, and support for AI almost everywhere

Women trust AI less than men and are less supportive of it in every application asked — from the mundane to the high-stakes.

Just 35.3% of women say they trust AI to do what is right ("somewhat" or "a lot"), versus 47.3% of men — a 12-point gap that survives full demographic controls (men +0.16 on the 1–4 trust scale, p <.001). The pattern extends to policy: across all eleven domains where the survey asked whether AI use should be supported, men are more supportive without exception. The gaps are largest for the highest-stakes, most physical-risk uses — national defense (48.7% of men vs. 30.3% of women support), self-driving vehicles (35.6% vs. 17.9%), semi-autonomous weapons (47.4% vs. 30.5%), cybersecurity (56.1% vs. 39.3%), and medical diagnosis (42.8% vs. 26.5%) — but even in low-support domains like AI in criminal sentencing or hiring, men run about ten points higher. Men are also more than twice as likely to support a local AI data center (31.6% vs. 15.3%).

One twist complicates a simple "men are AI boosters" reading: men are simultaneously more likely to agree that tech elites have too much influence over AI policy (59.4% vs. 50.2%). Greater enthusiasm coexists with sharper criticism of AI's power-holders — engagement, not uncritical approval.

What men and women actually use AI for

Men use AI more for nearly every purpose, and the gap is widest for technical work and — counter to a common assumption — for companionship.

Among AI users, men report more frequent use across essentially the entire task battery. At work, the gap is widest on the most technical tasks: writing or fixing code (mean 1.99 vs. 1.64 on a 1–4 scale) and data analysis (2.31 vs. 1.93), and narrowest — essentially nil — on writing emails (2.23 vs. 2.13). The tool data mirror this: men skew most heavily toward developer and frontier tools (Cursor and GitHub Copilot show some of the largest use gaps), while Grammarly is the one tool women use at the same rate or more. (Per-tool awareness comparisons — such as a male lead on Grok, DeepSeek, and Claude and a slight female edge on Grammarly — come from the ai_know battery flagged in Methods and are reported directionally only.)

The more surprising result is in personal and social use. A common assumption is that women are the heavier users of AI for companionship and emotional support; the data show the reverse. Men report more use of AI for casual conversations (57% vs. 43% any use), for emotional or romantic conversations (38% vs. 25%), for dating and relationship advice (42% vs. 32%), and they turn to AI chatbots for emotional support more frequently than women do. Among the minority who have used AI for emotional support, both genders rate it about equally helpful. The one clearly female-leaning use anywhere in the battery is rental-search assistance within the (small) group who have used AI for housing. Health and wellness is the most gender-neutral domain — men and women use AI for health advice, cooking, and nutrition at nearly equal rates. (One cross-survey caveat: Pew's June 2026 gender-gap report finds women slightly more likely to use chatbots for "emotional support or advice" — 11% vs. 8% — the opposite of the pattern here. The two are measuring adjacent but distinct things: the supplement asks about the frequency of casual and romantic "conversations" with AI, Pew about support-seeking as a chatbot use case. The male-leaning companionship result should therefore be read as specific to conversational and romantic framings, and treated as provisional.)

It is not the work they do

The most natural explanation for men's higher workplace AI use — that men hold more AI-amenable jobs — does not hold. The gap barely moves when the kind of work is held constant.

Men are more likely than women to do the kinds of tasks that pull AI into a job — data analysis, managing others, budget decisions — and to work in senior roles and remotely, all of which independently raise AI use. So a reasonable hypothesis is that the gender gap in workplace AI use is really an occupational gap in disguise. It is not. In a survey-weighted logistic model of "uses AI at work," men's odds are 1.45 times women's after controlling for the six standard demographics. Adding a full set of job-type controls — the six work-task flags, remote-work status, and organizational seniority — raises the model's fit substantially (McFadden R² 0.13 → 0.21), confirming those job features matter a great deal for AI use. But the male odds ratio is essentially unchanged: 1.46. Men and women who do the same kind of work, at the same seniority, in the same work arrangement, still diverge in whether they use AI — by almost exactly as much as before the controls.

That is a consequential null. It means the gender gap in AI use is not a byproduct of occupational sorting that will close on its own as women move into AI-adjacent roles. Something about propensity to adopt — plausibly the confidence gap documented above, plus the enthusiasm gap — operates on top of, and independently of, the work itself.

Where the gap is widest: it grows up the ladder

The gender gap is not largest among the disconnected; it is largest among the most engaged — the highly educated, the affluent, and the strongly partisan.

Because men lead in every subgroup — the gap never once reverses across age, income, education, race, urbanicity, or party — the question is where it is widest. The answer inverts the usual "digital divide" picture. The gap grows monotonically up the education and income ladders: in AI use at work it runs from +8 points among workers with a high-school education or less to +14 among graduate-degree holders; in trust it explodes from a 5-point gap among the least-educated to a 24-point gap among graduate-degree holders — the single largest gap in the entire analysis. It is likewise widest in the top income tercile (+18 on trust) and among Strong Republicans (+18), and narrowest among the least-educated, lowest-income, oldest (71–80), rural, and politically independent. By party the pattern is U-shaped: large at both partisan poles, smallest among pure independents.

The self-rated skill gap follows the same shape — widest among graduate-degree holders (a 0.52-point confidence gap) and Strong Republicans, near-parity among the oldest and least-educated. In other words, the environments where AI has most saturated — educated, affluent, professional — are precisely where men and women have pulled furthest apart.

At work: more upside for men, more caution from women

Men report more of AI's workplace benefits and more of its risky behaviors; women report more of its downsides and more caution.

The workplace items divide cleanly. On the upside, men are more likely to say AI increases their productivity (42% vs. 32%), improves their work quality (41% vs. 35%), and makes their job easier (48% vs. 41%), and far more likely to expect AI to make workers like them more productive in the long run (37% vs. 25%). They also perceive more organizational pressure to adopt AI on all four pressure items. On the downside, women are more likely to say AI's five-year impact on their own job will be "only negative" (20% vs. 14%), while men are more likely to say "only positive" (20% vs. 13%). Notably, on the blunt job-loss questions — is AI costing jobs like mine now, or ultimately — the gender gap nearly vanishes and slightly reverses (women marginally more likely to say yes, neither notable): women's greater pessimism is about their own trajectory, not aggregate displacement.

The clearest behavioral gap is caution. Asked what they have entered into workplace AI tools, 46% of women selected none of the sensitive categories, versus 33% of men; men are markedly more likely to have fed in financial data (41% vs. 29%), proprietary code or internal documents (30% vs. 19%), and client details (28% vs. 22%). Whether that reflects women's greater risk-aversion or men's greater exposure to data-rich tasks, the shadow-AI data-governance risk documented elsewhere in this series is disproportionately a male behavior. On disclosure itself, though, the genders are similar — there is little evidence women hide their AI use more or less than men.

The "not sure" gap

A recurring, easily-missed pattern: on AI-policy and knowledge questions, women are consistently more likely to answer "not sure" — so part of the measured gap is opinion withheld, not opposition.

Across the data-center energy question (women 15 points more likely to say "not sure"), the data-center impacts list, the value of degrees under AI, and who should teach AI skills, women disproportionately select "not sure" rather than taking a side. On the degree-value items, for instance, men and women are about equally likely to say AI makes a degree less valuable; women's distinctive response is elevated uncertainty (roughly 18–20% "not sure" vs. ~11% for men), not greater pessimism. This matters for interpretation. Some of the apparent attitudinal gender gap — especially on unfamiliar, technical policy topics — is women declining to offer an opinion on a subject they feel less informed about, which is itself downstream of the confidence gap. It is a caution against reading every gap as substantive disagreement, and a reminder that the confidence gap has measurement consequences well beyond self-rated skill.

Conclusion

Measured across nearly the entire surface of artificial intelligence, the gender gap is remarkable less for any single number than for its uniformity: men report more use, more intensity, more trust, more enthusiasm, and more support, on item after item, in subgroup after subgroup, and the gap never reverses. Three findings give that uniformity its shape. The skill gap is a gap in confidence, not competence — women underrate abilities they demonstrably have. The gap is not a product of the different jobs men and women hold; it persists undiminished when the work is held constant. And it is widest not among the disconnected but among the most educated, affluent, and engaged — the gap grows as AI's presence grows.

For anyone trying to widen AI's benefits, the distributional implication is pointed. If AI use compounds into productivity, wage, and career advantages, a gender gap that is largest among the highly educated and unexplained by occupation will tend to widen existing professional inequalities rather than being absorbed as women enter AI-adjacent roles. And because a meaningful part of the gap is confidence rather than capability, at least some of it is addressable — not by teaching women skills they already have, but by closing the gap between what they can do and what they believe they can do. This report maps the gap; the companion reports on trust, governance, and reskilling take up what to do about it.

Appendix A — Methods

Data source. AI@Work survey — CHIP50 Wave 38.6 (Civic Health and Institutions Project / COVID States Project). 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 remains a broad adult sample. Estimates are weighted to U.S. Census population targets. Data collection was ongoing at extract; figures are preliminary. The cross-technology comparison ("Is AI special?") draws social-platform adoption by gender from the same research program's core panel (weighted shares on the use_* columns) and is corroborated against Pew Research Center's platform-by-gender data.

Gender measure and bases. Gender is the vendor two-category variable (Male/Female). Each item is analyzed on its correct base as defined by the questionnaire's skip logic, and every figure states its base: all adults (unweighted n ≈ 23,300 with valid gender; ~10,960 men / ~12,470 women); the employed (≈ 12,400; ~6,920 men / ~5,485 women); AI users (≈ 17,500); or smaller filtered blocks — AI-using workers, those reporting an employer AI push, those who don't always disclose, respondents with hiring authority, business owners, students, and parents — each of which is labeled and, where small, flagged for reliability. Business-owner items rest on female bases of ~200–400 unweighted and are reported cautiously.

Metrics. For ordinal items the report uses either the weighted mean or a clearly-stated top-box share, whichever is more interpretable and least distorted by "not sure" codes; direction (including reverse-coded items such as ai_proficiency and the ai_use_support support/oppose scale) is stated wherever cited. For binary select-all batteries the metric is the weighted % selecting. The ai_freq_* tool battery is routed to varying subsets, so "% any use" is within each tool's asked subset, not population penetration. The ai_know awareness battery in the fielded instrument covers seventeen real, named tools (ChatGPT, Gemini, Claude, Grok, Copilot, DeepSeek, NotebookLM, Character.AI, Midjourney, Perplexity, Meta AI, Snapchat My AI, GitHub Copilot, Cursor, Grammarly, QuillBot, and Notion AI) and contains no fictitious catch-item. Note (data-integrity review): this battery is stored in a form the analysis tool mishandles — a gender crosstab of the ChatGPT item returns an impossible rate above 100% with a negative non-user count — so the by-gender awareness figures could not be independently reproduced and are reported with caution / directionally only; the headline near-parity in awareness is corroborated externally by Pew (identical ChatGPT use across genders). The substantive gender findings elsewhere in this report — work use, intensity, trust, self-rated and objective skill, purposes, affect, and the regressions — do not depend on the awareness battery and reconcile against live data.

Objective literacy. The 12-item ai_quiz battery is scored against the codebook answer key; "% correct" is the share choosing the keyed response per item, averaged across items. The ai_center_quiz (data-center facts) is scored the same way.

Intersections. The gender gap in three anchor outcomes — any AI use at work, trust in AI, and self-rated proficiency — was computed within each level of age (age_cat_8), income (income_cat_10 brackets grouped into terciles), education (education_cat), race (the race_* flags), community type (urban_type), and 7-point party (party7), by breaking each outcome down by gender within the filtered subgroup. Cells with unweighted n < 30 were flagged; none required suppression.

Mediation. Two survey-weighted logistic models of the binary "uses AI at work" (employed base; n ≈ 12,340–12,390) were compared: Model 1 with the six standard demographics (male OR 1.45, McFadden 0.131); Model 2 adding the six work_context task flags, remote_work, and indr_role_level seniority (male OR 1.46, McFadden 0.209). The near-identical male OR across the two is the mediation result. Two survey-weighted OLS models confirm the trust and confidence gaps survive the six demographics: pol_trust_ai male coefficient +0.163 (p <.001); ai_proficiency male coefficient −0.199 (p <.001, negative = more skilled). Standard errors are model-based, not design-based.

Significance discipline. A gap is called "surviving controls" only where the regression coefficient is significant at p <.05. All subgroup and crosstab comparisons are descriptive.

Reproducibility. All values were copied verbatim from the underlying tabulations. Item wording, coding, and skip logic were taken from the project's AI@Work survey codebook.

Appendix B — Data tables

The male odds ratio is unchanged by the job-type controls even though those controls are strong and lift model fit — men's higher AI use is not explained by the kind of work they do. * p <.001.