This post is the sixth in a series titled “AI and the Future of Work,” hosted on the Notre Dame-IBM Technology Ethics Lab’s blog. In this series, we will examine the implications of AI for jobs and society: how tasks and roles might change, how organizations can adapt, how workers are affected, and what governance frameworks are being considered to create and support beneficial and responsible outcomes.
The emerging consensus is clear: by 2030, roughly 60% of the global workforce will need reskilling—retraining workers in competencies such as data literacy, human-AI collaboration, and adaptive problem-solving that automated systems cannot readily replicate. Yet the institutions best positioned to fund this transition are underinvesting. The United States spends just 0.1% of GDP on active labor market policies (one-fifth the OECD average of roughly 0.5%), a gap equivalent to approximately $100 billion annually. Only 6% of firms are actively reskilling workers for AI, even as 85% pledge to prioritize upskilling (for more about reskilling vs upskilling, see our previous Blog). This gap between rhetoric and action—while AI reshapes an estimated 40% of global jobs—raises a fundamental ethical question: who bears the costs of reskilling while artificial intelligence is actively transforming the nature of work?
The World Economic Forum's 2025 Future of Jobs Report projects 170 million new roles created and 92 million displaced by 2030. The burden falls unevenly across lines of income, education, gender, geography, disability, and race, compounding existing inequalities rather than disrupting them. This blog examines the evidence across employer obligations, government responses, distributional impacts, educational adaptation, and ethical frameworks to inform an analytical assessment of cost distribution in AI-era reskilling.
Corporate investment lags far behind corporate rhetoric
International frameworks increasingly recognize employer responsibility for AI-driven workforce transitions, yet the gap between institutional expectations and corporate behavior remains vast. The EU AI Act's Article 4 established the world's first legally binding employer obligation for AI literacy, requiring all providers and deployers of AI systems to ensure staff has "a sufficient level of AI literacy." Similarly, the Organization for Economic Co-operation and Development (OECD), the International Labor Organization (ILO), the International Monetary Fund (IMF), and the World Economic Forum (WEF) all call for shared responsibility among employers, governments, and workers, but all emphasize that employers must play a significantly larger role than they currently do.
The data paints a stark picture. U.S. total training expenditures reached $102.8 billion in 2025, though nearly two-thirds of that figure — $64.7 billion — covers payroll for training staff rather than direct investment in learning content or delivery. Per-employee spending remains modest: the average organization allocates 2–4% of payroll to learning and development. McKinsey's analysis reveals that companies invest roughly 7.5 times more on technology than on the human capital needed to use it effectively. Among firms in the WEF's Global Lighthouse Network, for every $2 spent on technology, top performers spend $3 on process revamping and $5 on scaling and capability building. Most companies invert this ratio. Only 41% of employees report their company provides AI skills training, and 27% neither upskilling nor reskilling training in the past 5 years of their careers, with 65% of that group seeking training independently.
One idea—the beneficiary principle (the ethical argument that those who profit from a change should bear proportionate costs of that change)—finds strong empirical support. McKinsey sizes AI's productivity opportunity at $4.4 trillion in added growth potential, while PwC data shows workers with advanced AI skills commanding a 56% wage premium in 2024. The IMF projects that "in most scenarios, AI will likely worsen overall inequality," with gains accruing disproportionately to capital owners and high-skilled workers. Yet reskilling existing employees saves 70–92% compared to external hiring and delivers average first-year productivity gains of 10–15%, suggesting that the business case and the ethical case align.
The ILO's Termination of Employment Recommendation established an international norm that employers should explore alternatives to job loss during technological transitions, including retraining and consultation with workers. The ILO's more recent work frames the imperative sharply: "the transition to generative AI in the workplace must be managed and not left to chance," requiring "reskilling and adaptation support for workers in exposed roles."
Government programs are underfunded and outpaced by AI adoption
Public reskilling infrastructure faces three compounding challenges: insufficient funding, inadequate coverage, and a speed mismatch with the pace of AI adoption. In the United States, the Workforce Innovation and Opportunity Act (WIOA) provides roughly $2.9 billion annually for all 50 states and serves approximately 220,000 learners per year at about $2,000 per learner. Harvard's Project on Workforce describes this as a "low-resource, low-efficacy equilibrium." Since 2001, federal workforce development funding has declined approximately 50% in real terms. The WEF estimates U.S. reskilling costs for displaced workers alone at $34 billion, or $24,800 per worker.
Comparative data reveals dramatically different national approaches. In Singapore's SkillsFuture program, every citizen aged 25 and over receives a S$500 training credit, with mid-career workers aged 40+ receiving an additional S$4,000 top-up since 2024. Training participation rose from 35% to 50% of the resident labor force between 2015 and 2021, with more than 600,000 individuals taking up training with SkillsFuture support in 2025 alone. The program supports approximately 1,600 AI-related courses. South Korea committed 1.4 trillion won (~$960 million) in November 2025 to its first-ever national AI talent development blueprint, spanning elementary school through doctoral programs. The EU's European Social Fund+ allocates €142.7 billion for 2021–2027, with roughly one-third dedicated to education and skills, and the Commission launched a Skills Guarantee pilot in late 2025.
The speed mismatch is particularly concerning. AI use in the U.S. workplace jumped from 8% to 35% between 2023 and 2024; by spring 2025, 47% of workers used AI tools at least monthly. Meanwhile, WIOA has not been reauthorized since 2014, and expired in 2020. The EU AI Act's literacy requirements only took effect in February 2025, and the full EU Skills Guarantee won't be operational until 2028–2034. The IMF, OECD, ILO, and World Bank converge on a shared recommendation: dramatically increased public investment in reskilling, stronger social safety nets including unemployment and wage insurance, and tax reforms to redirect revenue toward human capital. The IMF specifically recommends against a direct "AI tax" as it is believed to stifle innovation, but urges strengthening capital income taxes and reconsidering corporate tax systems that inadvertently incentivize rapid displacement of human labor. The JUST Capital analysis highlights that the U.S. spends only 0.1% of GDP on active labor market policies, falling last among OECD nations, compared to Denmark's approximately 2%.
Inequality compounds at every level of the transition
AI-driven labor transitions do not affect all workers equally, and the evidence on distributional impacts is deeply troubling. The pattern is consistent: those most needing reskilling are systematically least likely to receive it.
Gender disparities are among the most pronounced. The ILO's 2025 refined global index finds that in high-income countries, 9.6% of female employment faces the highest automation risk, versus 3.5% for men: a 2.5-to-1 ratio driven by women's overrepresentation in clerical and administrative roles. Brookings' 2026 adaptive capacity analysis found that of the 6.1 million U.S. workers with both high AI exposure and low adaptive capacity, 86% are women. The digital gender divide exacerbates this: 71% of AI-skilled workers are men. In low-income countries, only 20% of women are connected to the internet.
Racial inequalities compound displacement risk. McKinsey's 2023 analysis found that 24% of Black workers hold positions in occupations with greater than 75% automation potential, versus 20% of white workers. Generative AI could widen the racial wealth gap by an estimated $43 billion annually by 2045, particularly by automating "gateway" jobs (positions paying above $42,000 that don't require college degrees) which represent critical pathways to middle-class incomes for workers without bachelor's degrees.
Age and education create additional fault lines. Only 24% of workers aged 55–65 participate in job-related training, versus 41% of those aged 45–54. A third of workers aged 55–65 have no computer skills or experience. Meanwhile, early-career workers face a different problem: the IMF found that workers aged 22–25 in the most AI-exposed occupations experienced a 13% relative decline in employment since ChatGPT's release.
The OECD's Skills Outlook 2025 documents a structural Matthew effect: adults whose parents obtained tertiary qualifications are more likely to engage in training that develops transferable, high-value skills, while those without tertiary-educated parents receive primarily job-specific or compliance-oriented training. Oxford Academic research confirms that firms' IT investments disproportionately benefit training for high-skilled workers, creating a positive feedback loop that widens within-firm wage inequality.
Disability creates yet another axis of exclusion. An estimated 1.3 billion people globally live with disabilities, most of working age, yet they face a persistent employment gap that AI threatens to widen from both ends. The ILO's Global Business and Disability Network finds that while AI-powered assistive technologies — real-time captioning, speech-to-text, and generative AI task support — can empower individual workers with disabilities, institutional uses of AI systematically disadvantage them: AI-powered HR tools filter out candidates with 'non-standard' facial expressions, speech patterns, or gaps in employment history, and AI systems trained on historical hiring data reproduce the underrepresentation of persons with disabilities in the workforce as if it were a desirable pattern. A 2025 study across 27 high-tech developed countries confirmed that AI increases unemployment among disabled individuals due to task automation and skill mismatches, with the effect only partially mitigated by advanced education. In the United States, only 22.5% of disabled working-age adults are in the labor force. Most reskilling and upskilling programs, meanwhile, are designed without the disability lens: inaccessible platforms, rigid scheduling, and failure to involve persons with disabilities in program design mean that the very workers most vulnerable to AI displacement are structurally excluded from the transition infrastructure meant to support them.
The Global South faces a compounding digital divide
Developing economies confront a qualitatively different set of challenges. The IMF estimates AI exposure at just 26% in low-income countries versus 60% in advanced economies. An IMF 2025 working paper projects that AI-driven growth in advanced economies could be more than double that in low-income countries, creating what researchers term an "inverse Balassa-Samuelson effect" that widens cross-country inequality.
The scale of the infrastructure gap is staggering: 2.6 billion people still lack internet access. In sub-Saharan Africa, fixed broadband costs average 20% of per capita gross national income, versus less than 1% in North America. Only 11% of university graduates in sub-Saharan Africa received formal digital training. Yet, the International Finance Corporation estimates 230 million African jobs will require digital skills by 2030.
More than 2 billion workers globally (58% of the workforce) remain in informal employment, where access to employer-provided training or state-supported skills development is minimal. The ILO warns of a white-collar bypass risk: clerical and administrative jobs that historically expanded the middle class and enabled women's workforce participation in developing countries may never grow at scale because AI can automate these tasks from the outset. This threatens to eliminate a critical rung on the development ladder.
The World Bank's 2025 analysis of East Asia and the Pacific found that robot adoption created approximately 2 million new jobs for skilled workers but displaced 1.4 million low-skilled formal workers between 2018 and 2022. Only about 10% of jobs in the region are complementary to AI, versus 30% in advanced economies. Basic generative AI licensing costs of $20–30 per user per month are prohibitive for small enterprises in informal sectors.
Education systems are adapting, but too slowly and too unevenly
An Inside Higher Ed survey found that only 9% of Chief Technology Officers believed higher education is prepared for AI's rise. Research indicates that 77% of new AI-related jobs require master's degrees or equivalent advanced training, far above the 35% education requirement for the jobs they're displacing.
Community colleges represent a critical but underresourced response. The National Applied AI Consortium, launched in 2024 with NSF support, now reaches 320 colleges across 46 states, training over 1,000 faculty and reaching 31,000 students. The California Community College system launched an AI Fellows Program in fall 2025. But offerings and specific trainings vary widely by institutional resources. As one community college leader noted, "If jobs start getting closed off only to graduates of colleges that are innovative enough to make it in this new economy, then we're going to see a reduction in career ladders."
Online platforms are scaling rapidly: Coursera reports 7.4 million AI enrollments in 2024. Micro-credentials show promising signals: 92% of employers say they're more likely to hire candidates with generative AI micro-credentials, and credential earners see a 34% higher job placement rate within six months. Yet these platforms largely serve already-educated, already-connected populations. The JFF found that leading AI literacy curricula are written at an 11th grade reading level, well above the proficiency of millions of adult learners.
Moreover, the credentialing model itself faces an emergent integrity crisis. Agentic AI tools can now complete entire online courses and credential assessments autonomously on a user's behalf. In early 2026, a tool called Einstein, designed to interface directly with the Canvas learning management system, demonstrated the capacity to watch lectures, write papers, participate in discussions, and submit assignments automatically, prompting the Modern Language Association to warn of “a fully automated loop in which assignments are generated by AI with the support of a learning management system, AI-generated content is submitted by an agentic AI on behalf of the student, and AI-driven metrics evaluate the work on behalf of the instructor.” Compliance training faces the same vulnerability: researchers found that ChatGPT in agent mode can be pointed at a training course URL and will navigate, watch videos, answer questions, and complete the course with nothing more than a few keystrokes. If micro-credentials can be earned without learning, their value as signals of human capital collapses and therefore undermining the very reskilling infrastructure these platforms are meant to provide.
Ethical frameworks converge on shared but unequal obligations
Multiple ethical traditions offer guidance on distributing reskilling costs, and while they differ in emphasis, they converge on a core insight: leaving the cost primarily to individual workers is unjust.
For example, Rawlsian justice applies the difference principle: AI-driven economic gains are justifiable only insofar as they benefit the least advantaged members of society. Behind a "veil of ignorance" — not knowing whether one would be a tech executive or a displaced clerical worker — rational actors would choose systems ensuring comprehensive reskilling funded by beneficiaries rather than victims. PNAS experimental research confirms that when people reason behind the veil, they adopt maximin (safety-first) strategies prioritizing the worst-off.
Another example is the "just transition" framework, originally developed for climate policy by labor leader Tony Mazzocchi and formalized in the ILO's 2015 guidelines, which is increasingly applied to AI. The ILO's April 2026 technical meeting explicitly grouped "decent work, productivity and a just transition arising from artificial intelligence." The framework's five principles — equity, inclusion, shared responsibility, social dialogue, and advance planning — offer a practical governance template. The WEF's 2024 Davos discussions explicitly connected just transition language from climate to AI.
Final Thoughts
The central tension in AI reskilling is not whether the cost is worth bearing; the economic returns are well-documented, from the GI Bill's $6.90 return per dollar invested to McKinsey's evidence that human-capital-intensive technology adoption outperforms alternatives. The tension is structural: those capturing AI's enormous productivity gains are not the same populations bearing the displacement costs. The beneficiary principle, Rawlsian justice, and the just transition framework all point toward a distribution of costs proportionate to benefits received and capacities held.
What distinguishes the current moment from previous technological transitions is both scale and speed. The WEF projects 22% of all jobs disrupted by 2030. AI workplace adoption jumped from 8% to 35% in a single year. Yet institutional responses operate at a fundamentally different tempo. The window for proactive investment is narrowing.
Employers, governments, and educational institutions each hold distinct but overlapping obligations, shaped by proximity to the disruption, capacity to act, and share of the gains. The same AI systems that generate trillion-dollar productivity estimates can either widen the gap between those who benefit and those who adjust, or become the occasion for reinvestment in human capability at a scale that matches the transformation underway. What determines which outcome prevails is not the technology itself, but whether the institutions shaping its adoption treat reskilling as a cost to be minimized or as an obligation to workers and to the long-term sustainability of the economic systems AI is reshaping.
Perla Khattar is a J.S.D. candidate at the University of Notre Dame Law School and a Research Fellow at the Notre Dame-IBM Tech Ethics Lab. In her research, she analyzes emerging technologies such as synthetic data, neurodata, quantum computing, and AI through a privacy lens, focusing on how law and policy can anticipate technological disruption while safeguarding individual rights.