When AI Enters the Workplace, Who Faces Greater Risks? A Gendered Analysis

AI summary

Overview

Research area: AI safety and ethics / labour economics, specifically the gendered distribution of AI exposure across occupations.

Technical level: Intermediate. The statistical machinery is standard (ANOVA, Kruskal-Wallis, Tukey and Dunn post-hoc tests), but the paper assumes familiarity with occupational classification systems (SOC, OCC, O*NET-SOC), two distinct AI-exposure indices, and a purpose-built composite "SW score" for skill and wage levels.

Scope in one sentence: The paper links O*NET occupational data, U.S. Census earnings and workforce counts, and two AI-exposure indices to test whether AI exposure falls differently across the skill and wage distribution in male-dominated versus female-dominated occupations.

What This Paper Is About

Prior work has largely asked whether female-dominated occupations are more or less exposed to AI than male-dominated ones, but has not asked where within those occupations the exposure sits along the skill and pay ladder. That matters because the difference between AI arriving as a productivity aid to a well-paid professional and AI arriving as task automation in a low-paid clerical job is exactly what determines whether women gain or lose. The paper builds a new linked dataset that answers that finer-grained question and then triangulates the statistical results against existing labour-market literature to interpret what the exposure patterns likely mean.

Key Contributions

1. A novel linked dataset. The authors join O*NET occupational characteristics, 2024 American Community Survey gender-disaggregated employment counts and median earnings, and two AI-exposure indices, producing analytical samples of 505 occupations for the AII and 394 occupations for the Anthropic Index.

2. A dual-index design. Rather than relying on a single exposure measure, the paper contrasts the Anthropic Index (LLM-related exposure, derived from Claude usage data) with the AI Impact Index or AII (broader AI innovation, derived from patent–task similarity), capturing two distinct technological paradigms.

3. The SW score. A composite measure combining log-transformed median earnings and O*NET job zone training/education requirements into z-scores, weighted w = 0.6 toward earnings, to place each occupation on a combined skill-and-wage axis.

4. The Geometric Exposure Index (GEI). A new aggregation method for mapping multiple SOC occupations onto a single OCC code. It multiplies prevalence (the probability the index is greater than zero) by the geometric mean of positive values only, giving an expected-value interpretation while handling the right-skewed, zero-inflated distributions of both indices.

Main Findings

• Female-dominated occupations show higher LLM exposure. Occupations with more than 60% female workers have higher Anthropic Index values than occupations with fewer than 40% female workers (74.3 vs. 24.3). The pattern is non-monotonic: mixed occupations (40–59% female) show the highest values, male-dominated the lowest, female-dominated in between. ANOVA confirmed significant differences (F = 18.39, p < 0.001), and Tukey pairwise comparisons showed male-dominated occupations significantly lower than both mixed and female-dominated occupations (p < 0.001).

• Male-dominated occupations show higher broader AI exposure. For the AII, male-dominated occupations have the highest values, followed by mixed, with female-dominated lowest (77.73 vs. 62.05). Kruskal-Wallis confirmed significant differences (H = 7.99, p < 0.05), and Dunn's post-hoc test showed a significant male-dominated versus female-dominated difference (p < 0.05).

• Male-dominated occupations concentrate exposure at the top of the skill and wage ladder. Within male-dominated occupations, statistically significant differences in exposure appear across skill and wage levels for both the Anthropic Index (reported as F = 16.18, p > 0.01) and the AII (F = 8.35, p < 0.05), indicating concentration in higher-skilled, higher-paid roles.

• Female-dominated occupations spread exposure evenly. Within female-dominated occupations, exposure is distributed more uniformly across the skill and wage spectrum, and the statistical tests were not significant for either index (p > 0.05).

• Occupational segregation is visible in the raw data. The most male-dominated occupations listed include elevator and escalator installers and repairers (0.6% female, 99.4% male, 27,775 workers) and earth drillers (1.3% female, 98.7% male, 23,261 workers). The most female-dominated include skincare specialists (98.3% female, 52,438 workers) and speech-language pathologists (95.3% female, 139,011 workers).

• Low-SW female-dominated occupations cluster in administrative and service work. The authors map their lowest SW bins against occupation lists in the Jobs and Skills Australia report, the World Economic Forum Future of Jobs analysis, and ILO evidence on generative AI exposure, finding strong overlap with roles the literature identifies as susceptible to routine task automation and restructuring.

• Gender gaps appear in the underlying wage data. In the illustrative Census sample, chief executives show estimated median earnings of $191,756 for men versus $151,010 for women; financial managers show $124,972 versus $82,680.

• Adoption and benefit gaps are documented in prior literature. The review cites survey evidence that men report higher generative AI use than women (50% versus 37%), with knowledge differences accounting for roughly three-quarters of the gap, and cross-country evidence from 18 studies covering more than 140,000 individuals showing consistently lower female adoption of generative AI tools.

Methodology in Plain English

The researchers assembled four data sources and stitched them together by occupation code.

O*NET, maintained by the U.S. Department of Labor, describes roughly 1,000 occupations with their tasks, importance scores, and job zones (five levels reflecting required education, training, and experience, from zone 1 to zone 5). The 2024 one-year American Community Survey provides the number of men and women in each detailed occupation (tables B24124, B24125, B24126) and median annual earnings by gender (tables B24121, B24122, B24123, with B24124/B24125/B24126 used for workforce counts and B24121/B24122/B24123 for earnings). Two exposure indices come from prior work: the AII, which measures the share of an occupation's tasks whose text is highly similar to AI-related patent text, with exposure flagged using the 90th percentile similarity threshold; and the Anthropic Index, an observed-exposure score built from Claude usage data from August 2025 and November 2025, comprising 2 million Claude.ai observations and 2 million first-party API observations, with task coverage requiring a minimum of 100 work-related uses, equal to 0.0025% of total traffic.

Because the Census uses OCC codes, exposure indices use O*NET-SOC codes, and everything must pass through SOC, the authors perform two mapping steps. Some OCC codes map to several SOC codes, so the GEI aggregates them. Occupations where the standard deviation of the constituent AII values exceeded 0.15 were excluded to limit aggregation noise (five occupations), a threshold the authors report results are robust to at alternative cutoffs of 0.10 or 0.20.

Coverage was uneven. The AII was available for 873 of 923 ONET occupations, and 517 of 570 OCC occupations were matched, yielding a final dataset of 505 occupations retaining 86.1% of detailed Census occupations, with coverage above 80% in nearly all sectors and 100% in several. Transportation was the notable shortfall at 63.6%, and Military Specific was entirely absent. The Anthropic Index was available for 756 of 923 ONET occupations, matched to 414 of 570 OCC occupations; the final sample retained 394 of 570 Census occupations after 20 matched occupations were dropped for missing gender-disaggregated data. Anthropic coverage was at least 50% in every sector except Military Specific.

With the integrated dataset, the authors ran two analyses. The first grouped occupations by female share (under 40%, 40–59%, 60% and above) and compared exposure distributions. The second used the SW score to test whether exposure varies with skill and wage within male- and female-dominated occupations separately. Because the final step is interpretive — exposure is not the same as displacement — they compared the occupations landing in the lowest and highest SW bins against occupation lists from prior automation-exposure studies, using automated text matching (exact matching and Levenshtein distance) followed by manual review.

Why This Matters

Impact on research. The paper shifts the unit of analysis from "are female-dominated occupations exposed?" to "where within female-dominated occupations is exposure located?" It also introduces the GEI as a reusable aggregation method for skewed, zero-inflated exposure indices, and demonstrates that the choice of exposure index materially changes the answer — LLM-based and patent-based measures point in opposite directions on gender.

Real-world applications:

• Targeted reskilling programmes. The finding that low-SW female-dominated occupations combine higher automation-linked exposure with lower participation in adult learning (the paper cites OECD evidence on low-skilled workers) identifies a specific population for intervention rather than a broad sector.
• HR and workforce planning. Employers deciding where to deploy AI tools can distinguish between roles where AI complements high-skilled work and roles where it substitutes for tasks, and account for who currently holds each type of role.
• AI governance and audit. The cited evidence that identical salary-negotiation prompts produced lower recommended salaries for women illustrates why advisory and evaluative AI systems used in pay and progression decisions warrant review.
• Policy measurement. The paper shows that the headline conclusion about gender and AI exposure depends on which index is used, which is directly relevant to how governments and international bodies construct labour-market impact assessments.

Industry relevance. Organisations in administrative, clerical, and service sectors — where the paper finds female-dominated, low-SW roles concentrated — face the most direct workforce implications. AI vendors and platform operators have a stake because the Anthropic Index reflects actual Claude usage, meaning adoption patterns on a single platform shape measured exposure. Regulators and unions involved in automation-related bargaining also have direct use for the skill-and-wage breakdown, which is more actionable than an occupation-level average.

Future Directions

1. Extend the observation window for LLM usage data. The authors note the Anthropic Index rests on a relatively short observation window, which may not capture stable or representative patterns of LLM use across occupations, and that different LLMs may show different adoption and use patterns.

2. Harmonise coverage across exposure indices. The two indices do not cover identical occupation sets, which constrains direct comparability between LLM-specific exposure and broader AI innovation. Closing that gap would allow cleaner claims about distinct technological paradigms.

3. Move from exposure to realised outcomes. Both indices measure potential exposure, not whether AI adoption actually produces augmentation, automation, productivity gains, wage growth, or displacement. Linking exposure measures to observed wage, employment, and progression outcomes would test the risks the paper infers from literature triangulation.

4. Interpret the low-SW bins in more depth. The paper identifies that female-dominated low-SW occupations concentrate in administrative and service sectors and overlap with roles prior studies flag as automation-exposed, but the analysis is interpretive. Deeper occupational-level analysis of those specific roles, and of the training access available to them, is a natural extension.

Target Audience

Labour economists and AI-and-society researchers will find the dual-index design and the GEI method most useful. Policymakers and analysts working on gender equality, workforce transition, and AI regulation — including the international bodies whose reports the paper triangulates against — are the primary applied audience. HR, workforce planning, and responsible-AI practitioners in administrative and service sectors will benefit from the skill-and-wage breakdown even without following the statistical detail. Readers need comfort with regression-style group comparisons but not with machine learning methods, as the exposure indices are used as given rather than constructed here.

Authors’ abstract

Gender inequality remains a persistent structural feature of the labour market, shaping women's lifetime earnings and economic security. As artificial intelligence (AI) transforms organisational practices, there is growing concern that existing disparities may be unintentionally amplified through task automation, unequal access to upskilling opportunities, and differential returns obtained from technological change. In this paper, we examine how exposure to AI-driven innovation varies across male- and female-dominated occupations, with particular attention to differences across the skill and wage distribution. Using a novel dataset that links occupational characteristics to measures of AI exposure, we analyse how recent advances in Large Language Models (LLMs) and broader AI technologies are distributed across the labour market. Our findings show that, while AI exposure is generally concentrated in higher-skilled and higher-paid occupations for male-dominated occupations, female-dominated occupations display relatively uniform levels of exposure across both high-skilled, high-paid, and low-skilled, low-paid occupations. Moreover, we find that LLM-related exposure is higher in female-dominated occupations, while exposure to broader AI innovation remains more concentrated in male-dominated occupations. A triangulation of these results with existing literature suggests that women, particularly those in the most vulnerable positions (lower-skilled and lower-paid female-dominated occupations), may face greater exposure to forms of AI associated with task automation, job restructuring, reduction of wages and limited career progression.