1. Introduction: The Imperative for Understanding AI's Global Impact
The rapid advancement of artificial intelligence technologies has raised urgent questions about their impact on labor markets, a concern most pressing in developing countries where technological disruption intersects with profound existing vulnerabilities. Understanding which workers and occupations face the greatest exposure to AI-driven change is a critical research priority for crafting effective policy.
Measuring AI exposure presents unique challenges. Unlike prior automation waves targeting specific physical or routine cognitive tasks, modern AI, especially large language models (LLMs), has capabilities spanning multiple domains, potentially augmenting or substituting a broad range of human activities. This requires new analytical frameworks that account for nuanced relationships between AI capabilities and occupational tasks, while considering the vast differences in economic contexts, technological infrastructure, and labor market structures globally.
The stakes are illustrated starkly by Brazil. In 2023, 27.6% of Brazilian households experienced food insecurity, with 4.1% (over 3.2 million households) suffering severe food insecurity and hunger. Critically, the demographic profiles of these households—headed by women, non-white individuals, those with less education, and larger families—significantly overlap with those identified as highly exposed to AI-driven job displacement. This suggests AI-driven labor disruption could exacerbate existing inequalities, pushing vulnerable populations deeper into poverty and hunger. This paper addresses the urgent need for better measurement tools by examining the methodological landscape for assessing AI work exposure, analyzing their strengths and limitations in developing countries, and proposing a new open-source methodology to overcome current barriers. Accurate, contextually appropriate measurement is essential for policymakers to design interventions that protect vulnerable populations while harnessing AI’s potential benefits.
2. Foundational Methodologies: Building the Measurement Infrastructure
The measurement of AI exposure has evolved through increasingly sophisticated methodologies, each building on prior insights. Understanding this evolution is crucial for appreciating the strengths and limitations of current approaches, especially in the context of developing countries.
2.1 Webb (2020): The Patent-Based Pioneer
A foundational approach comes from Webb (2020), who created an automated system linking technological capabilities in patents to occupational tasks in databases like O*NET. The process begins by identifying AI-specific patents using keywords (e.g., "neural network"). It then extracts verb-noun pairs (e.g., "diagnose disease") from patent titles and, in parallel, from O*NET task descriptions. A key innovation is using WordNet to group semantically similar nouns, bridging the gap between specific technical language and broader occupational terms. The final exposure score for an occupation is a weighted average of the frequency of these verb-noun pairs in the AI patent corpus, reflecting how intensively AI innovation targets an occupation's tasks. This method is objective, replicable, and forward-looking. However, it measures technological targeting, not adoption, and may miss innovations in open-source AI or introduce noise through its semantic matching, limitations that are particularly acute in developing countries.
2.2 Brynjolfsson et al. (2023): The Expert Judgment Revolution
Brynjolfsson et al. (2023) advanced the field with their Suitability for Machine Learning (SML) index, which incorporates expert human judgment. Their methodology evaluates each Detailed Work Activity (DWA) in O*NET against a 23-question rubric based on the technical feasibility of applying machine learning. A key conceptual innovation is distinguishing between destructive digitalization (labor displacement, measured by the mean SML score) and transformative digitalization (job reconfiguration, measured by the standard deviation of SML scores). This reveals that many jobs face transformation rather than elimination. The approach's main limitations are that its crowdsourced evaluators may have biases reflecting developed-country contexts, and its focus on technical feasibility may overestimate real-world exposure by ignoring economic, legal, and social barriers to adoption.
2.3 Felten et al. (2023): The Comprehensive Framework
The Felten et al. (2023) methodology created a comprehensive framework linking ten distinct AI applications (e.g., image recognition, language modeling) to 52 human abilities from ONET. A crowd-sourced matrix, created by computer science experts, rates the relatedness between each AI application and human ability. An occupation's exposure score (AIOE) is then calculated by combining this matrix with ONET data on the importance of those abilities for the job. A crucial advantage is its adaptability; researchers can isolate exposure to a specific AI application, like language models, by adjusting the weighting. However, it relies on expert assessments that may not reflect on-the-ground work realities and remains agnostic about whether AI's impact will be substitutive or complementary.
2.4 Eloundou et al. (2023): The Task-Based Precision Approach
Representing the current frontier, Eloundou et al. (2023) developed a framework to assess LLM exposure by evaluating if a task's completion time could be reduced by at least 50%. Their rubric distinguishes between direct exposure (E1) (achievable with current tools) and complementary exposure (E2) (requiring future innovation). A key innovation is using both human annotators and GPT-4 for the evaluation, finding high correlation. The approach provides nuanced, practical insights, such as finding that female workers face disproportionately higher exposure. Its limitations include a focus on time reduction that may miss qualitative job changes and potential overestimation of exposure where social or regulatory norms require human oversight, a particularly relevant factor in diverse cultural contexts.
2.5 Pizzinelli et al. (2023): The Complementarity Revolution
Pizzinelli et al. (2023) revolutionized existing measures by introducing a complementarity adjustment. They recognized that high exposure could mean either substitution or augmentation. Their complementarity index (θ) uses O*NET data on factors like the criticality of decisions and consequences of error to assess whether society would allow AI to operate without human supervision. The resulting complementarity-adjusted AIOE (C-AIOE) fundamentally reframes vulnerability. Managers and professionals, despite high AIOE scores, have their adjusted exposure reduced due to high complementarity. Conversely, clerical and support workers emerge as most vulnerable. The framework’s main limitation is its reliance on current social and regulatory norms, which are likely to evolve as AI advances.
3. Applications to Developing Countries: Bridging the Measurement Gap
The foundational methodologies, developed primarily with U.S. data, face significant challenges when applied to developing countries, where work structures and technology adoption differ substantially.
3.1 Basic Adaptations: The Direct Application Challenge
The most common approach is the direct application of U.S.-based scores to other countries’ occupational structures, as seen in studies by Cazzaniga et al. (2024) and the OECD (2024). This method provides rapid global comparisons but suffers from a fundamental flaw: it ignores substantial within-occupation task differences across countries. An accountant in Ohio and one in Nigeria may share a job title but perform very different tasks, leading to systematic misestimation of AI's true impact.
3.2 Semantic Matching Innovations: The Carbonero Breakthrough
Addressing this, Carbonero et al. (2023) developed a semantic matching system linking U.S.-based exposure scores to individual worker skills reported in surveys from Lao PDR and Vietnam. This individual-level analysis revealed striking differences missed by direct application, showing exposure is far more heterogeneous in urban Vietnam than in agrarian Lao PDR. The approach is powerful but limited by its need for high-quality, individual-level skill-use data, which is rare in developing countries.
3.3 Advanced Cross-Country Frameworks: The Lewandowski Revolution
Lewandowski et al. (2025) achieved a breakthrough by adapting the Felten AIOE methodology to 50 countries, systematically addressing task composition differences. They created a mapping system that predicts O*NET ability levels from internationally comparable survey questions. Their groundbreaking finding is that within-occupation task differences account for roughly half of the global AI exposure divide. Workers in developing countries perform fundamentally different, less AI-exposed tasks within nominally similar occupations, suggesting direct application of U.S. measures systematically overestimates AI impact in the developing world.
3.4 Digital Divide Adjustments: The Gmyrek Innovation
Gmyrek et al. (2024) introduced another crucial layer of realism by creating a digital access adjustment. Recognizing that theoretical exposure is meaningless without infrastructure, they modified exposure scores for Latin America based on workers' actual access to digital technologies. Their sobering finding is that nearly half of all jobs that could be augmented by AI are held by workers hampered by inadequate digital access. This digital divide acts as both a buffer against displacement and a bottleneck to productivity.
3.5 Comparative Methodological Analysis: The EganadelSol Innovation
EganadelSol and BravoOrtega (2025) applied both the Felten and Webb methodologies to four Latin American countries, showing how methodological choice shapes results. Felten's approach identified higher-skilled workers as more exposed, while Webb's showed more varied patterns. This divergence is critical, as it suggests different policy priorities depending on the measurement tool used, and underscores that methodological choice is not neutral in identifying who is at risk.
3.6 Temporal Innovation in AI Assessment: The Benítez and Parrado Framework
Benítez and Parrado (2024) pioneered a dynamic approach, using "synthetic AI surveys" to assess occupational vulnerability across one-, five-, and ten-year horizons. Their Generated Index of Occupational Exposure (GENOE) shows a non-linear acceleration of risk over time. While innovative, the framework's reliance on an AI predicting its own future development is inherently speculative and may miss complex social or institutional adoption barriers.
3.7 Empirical Validation: The Ajimoti Real-World Impact Assessment
Moving from theory to validation, Ajimoti et al. (2025) linked real-world AI usage patterns in Brazil to actual employment outcomes from 2021-2024. After controlling for economic cycles, they found no statistically significant evidence of net job displacement from AI automation. This crucial empirical work suggests that theoretical exposure measures may overestimate the near-term pace of AI impact, likely due to low adoption rates and infrastructural gaps. However, the study's conclusions are preliminary, given its reliance on a single AI platform and a short time frame.
4. A Proposed Open-Source Framework for Measuring AI Exposure
Drawing from this review, we propose a novel framework centered on creating and publicly releasing a transparent, replicable, and contemporary toolkit for measuring AI work exposure. The urgent need for such a tool is underscored by Brazil's convergence of food insecurity and potential AI-driven labor displacement. Our primary goal is to address the gap left by existing indices that are often dated, opaque, or inaccessible.
We adopt and modernize the patent-based approach of Webb (2020), updating his AI patent corpus with keywords related to the recent generative-AI surge. This ensures the measure reflects the current technological frontier. We will then replicate the process of extracting verb-noun pairs, performing semantic harmonization, and calculating task- and occupation-level scores. Crucially, this work will deliver two public resources: open-source Python code to replicate the procedure and a complete dataset of updated AI-exposure scores. This toolkit is intended as a public good to empower the global research community.
This phase delivers two public resources:
- Open-source Python code with detailed documentation and worked examples that replicate the entire procedure;
- A complete dataset of updated AI-exposure scores for all O*NET tasks and occupations, accompanied by metadata on the patent corpus and calculation methods.
The final results and data can be accessed here.
5. The Imperative for Research: AI's Potential Impact on Severe Food Insecurity
AI's emergence as a General-Purpose Technology occurs in a world of profound deprivation, creating an imperative to understand its impact on basic human welfare, particularly food security.
The relationship is dual-natured. AI shows promise in agriculture, with applications enhancing crop yields and efficiency (HIGH TECH..., 2024; STONES, MALLERET and PINNOW, 2024). However, it also risks displacing agricultural labor and exacerbating inequality if its benefits are captured only by large, capitalized farms (FOOD AND AGRICULTURE ORGANIZATION OF THE UNITED NATIONS, 2022).
Beyond agriculture, AI's most significant effects on food security will likely be indirect, channeled through widespread labor market disruption that impacts household income (CAZZANIGA et al., 2024). Brazil presents an urgent case. The country's high rates of food insecurity (CIRÍACO et al., 2025) affect demographic groups—households headed by women, non-white individuals, and those with less education—that are also identified as highly exposed to AI automation (GMYREK; WINKLER; GARGANTA, 2024). This convergence of vulnerabilities could create a catastrophe where technological progress systematically undermines the economic stability of the most food-insecure households.
This is compounded by developing-country challenges, including a lack of granular data on what workers actually do (WORLD BANK, 2024), a persistent digital divide, and large informal sectors that offer no safety net for displaced workers. This confluence of factors creates an urgent need for research. Without accurate measurement of AI exposure, policymakers cannot effectively target interventions. The framework proposed here is an essential first step in providing the evidence needed to ensure technological progress reduces, rather than deepens, human suffering.
6. Conclusion and Future Directions
This review reveals both remarkable innovation and persistent gaps in AI work exposure measurement. Methodologies developed in and for advanced economies are fundamentally limited when applied to developing countries, where different task compositions and adoption patterns prevail. The finding by Lewandowski et al. (2025) that task differences account for half the global exposure divide underscores the danger of applying these measures without careful adaptation.
The Brazil case study illustrates the urgent human consequences, with a potential convergence of technological and nutritional vulnerabilities. The open-source framework proposed here is a step toward democratizing AI impact assessment and enabling more contextually appropriate measurement.
Future research should prioritize real-time monitoring systems (building on Ajimoti et al., 2025), standardized protocols for task-level data collection in developing countries, and the integration of exposure measures with broader welfare indicators like food security. The intersection of AI and food security deserves particular attention. Methodologically, dynamic frameworks that adapt to evolving AI capabilities are needed.
The urgency cannot be overstated. As AI advances, the window for proactive policy is closing. The measurement tools developed today will determine whether policymakers can harness AI's benefits while protecting the vulnerable. This global challenge demands international cooperation and transparent, collaborative research that places human welfare at the center of technological progress.
REFERENCES
AJIMOTI, B.; TOMAZ, V.; MAGED, H.; ABDULFATAH, M. S. A. Evaluating the risk of job displacement by transformative AI automation in developing countries: a case study on Brazil. [S.l.]: Apart Research, 28 Apr. 2025. [online]. Available at: https://apartresearch.com/project/evaluating-the-risk-of-job-displacement-by-transformative-ai-automation-in-developing-countries-a-case-study-on-brazil-829h. Accessed on: 1 Jul. 2025.
BENÍTEZ-RUEDA, M.; PARRADO, E. Mirror, mirror on the wall: which jobs will AI replace after all?: a new index of occupational exposure. Washington, DC, USA: Inter-American Development Bank, Department of Research and Chief Economist, Aug. 2024. IDB Working Paper Series, no. IDB-WP-1624. [online]. Available at: http://dx.doi.org/10.18235/0013125. Accessed on: 1 Jul. 2025.
BRYNJOLFSSON, E.; LI, D.; RAYMOND, L. R. Generative AI at work. Cambridge, MA, USA: National Bureau of Economic Research, Apr. 2023; rev. Nov. 2023. NBER Working Paper Series, no. 31161. [online]. Available at: http://www.nber.org/papers/w31161. Accessed on: 1 Jul. 2025.
CARBONERO, F.; DAVIES, J.; ERNST, E.; FOSSEN, F. M.; SAMAAN, D.; SORGNER, A. The impact of artificial intelligence on labor markets in developing countries: a new method with an illustration for Lao PDR and urban Viet Nam. Journal of Evolutionary Economics, v. 33, p. 707–736, Jul. 2023. [online]. Available at: https://doi.org/10.1007/s00191-023-00809-7. Accessed on: 1 Jul. 2025.
CAZZANIGA, M.; JAUMOTTE, F.; LI, L.; MELINA, G.; PANTON, A. J.; PIZZINELLI, C.; ROCKALL, E.; TAVARES, M. M. Gen-AI: Artificial Intelligence and the Future of Work. Washington, DC: International Monetary Fund, Jan. 2024. IMF Staff Discussion Note, no. SDN/2024/001. ISBN 979-8-40026-254-8. [online]. Available at: https://www.imf.org/-/media/Files/Publications/SDN/2024/English/SDNEA2024001.ashx. Accessed on: 1 Jul. 2025.
CIRÍACO, Juliane da Silva et al. Insegurança alimentar e nutricional no Brasil: análise de fatores determinantes em domicílios urbanos e rurais em 2023. Rio de Janeiro: Instituto de Pesquisa Econômica Aplicada (Ipea), May 2025. Texto para Discussão, no. 3121. 32 p.: il. [online]. Available at: https://repositorio.ipea.gov.br/bitstream/11058/17233/1/TD_3121_web.pdf. Accessed on: 1 Jul. 2025.
EGANA-DEL SOL, P.; BRAVO-ORTEGA, C. Artificial Intelligence and Labor Market Transformations in Latin America. Bonn: IZA Institute of Labor Economics, Feb. 2025. IZA Discussion Paper Series, no. 17746. ISSN 2365-9793. [online]. Available at: https://docs.iza.org/dp17746.pdf. Accessed on: 1 Jul. 2025.
ELOUNDOU, T.; MANNING, S.; MISHKIN, P.; ROCK, D. GPTs are GPTs: An early look at the labor market impact potential of large language models. arXiv preprint arXiv:2303.10130v5 [econ.GN], 22 Aug. 2023. [online]. Available at: https://arxiv.org/abs/2303.10130v5. Accessed on: 1 Jul. 2025.
FELTEN, E.; RAJ, M.; SEAMANS, R. How will language modelers like ChatGPT affect occupations and industries? [S.l.]: SSRN, 18 Mar. 2023. [online]. Available at: https://ssrn.com/abstract=4375268. Accessed on: 1 Jul. 2025.
FOOD AND AGRICULTURE ORGANIZATION OF THE UNITED NATIONS (FAO). The State of Food and Agriculture 2022: Leveraging automation in agriculture for transforming agrifood systems. Rome: Food and Agriculture Organization of the United Nations, 2022. [online]. Available at: https://doi.org/10.4060/cb9479en. Accessed on: 1 Jul. 2025.
GMYREK, P.; BERG, J.; BESCOND, D. Generative AI and jobs: A global analysis of potential effects on job quantity and quality. ILO Working Paper, no. 96. Geneva: International Labour Organization, Aug. 2023. [online]. Available at: https://doi.org/10.54394/FHEM8239. Accessed on: 1 Jul. 2025.
GMYREK, P.; WINKLER, H.; GARGANTA, S. Buffer or Bottleneck? Employment Exposure to Generative AI and the Digital Divide in Latin America. ILO Working Paper, no. 121. Geneva: International Labour Organization and The World Bank, Jul. 2024. [online]. Available at: https://doi.org/10.54394/TFZY7681. Accessed on: 1 Jul. 2025.
HIGH TECH, HIGH YIELDS? The Kenyan farmers deploying AI to increase productivity. The Guardian, London, 30 Sep. 2024. Global development. [online]. Available at: https://www.theguardian.com/world/2024/sep/30/high-tech-high-yields-the-kenyan-farmers-deploying-ai-to-increase-productivity. Accessed on: 1 Jul. 2025.
LEWANDOWSKI, P.; MADOŃ, K.; PARK, A. Workers’ Exposure to AI Across Development. IBS Working Paper, no. 02/2025. Warsaw: Institute for Structural Research (IBS), June 2025. [online]. Available at: https://ibs.org.pl/wp-content/uploads/2025/03/Workers_AI_exposure_across_development_IBS_WP_202502.pdf. Accessed on: 1 Jul. 2025.
MANNING, Sam. Addressing the U.S. Labor Market Impacts of Advanced AI. In response to the U.S. AI Action Plan RFI, Mar. 14, 2025. [online]. Available at: https://www.federalregister.gov/documents/2025/02/06/2025-02305/request-for-information-on-the-development-of-an-artificial-intelligence-ai-action-plan. Accessed on: 1 Jul. 2025.
MEIRELES, Thiago. Dessa vez é diferente? In: Panorama Setorial da Internet, São Paulo, vol. 16, no. 4, pp. 1–6, Nov. 2024. [online]. Available at: https://cetic.br/media/docs/publicacoes/6/20241218183020/ano-xvi-n-4-ia-mercado-trabalho.pdf. Accessed on: 1 Jul. 2025.
MONTAGNER, Paula. A Inteligência Artificial e o mercado de trabalho no Brasil: oportunidades, desafios e diretrizes para políticas públicas. In: Panorama Setorial da Internet, São Paulo, vol. 16, no. 4, pp. 23–26, Nov. 2024. [online]. Available at: https://cetic.br/media/docs/publicacoes/6/20241218183020/ano-xvi-n-4-ia-mercado-trabalho.pdf. Accessed on: 1 Jul. 2025.
OECD. Job Creation and Local Economic Development 2024: The Geography of Generative AI. Paris: OECD Publishing, 28 Nov. 2024. [online]. Available at: https://doi.org/10.1787/83325127-en. Accessed on: 1 Jul. 2025.
PIZZINELLI, C.; PANTON, A.; TAVARES, M. M.; CAZZANIGA, M.; LI, L. Labor Market Exposure to AI: Cross-country Differences and Distributional Implications. Washington, DC: International Monetary Fund, Oct. 2023. IMF Working Paper Series, no. 23/216. [online]. Available at: https://www.elibrary.imf.org/downloadpdf/view/journals/001/2023/216/001.2023.issue-216-en.pdf. Accessed on: 1 Jul. 2025.
STONES, L; MALLERET, C; PINNOW, F. Low-carbon milk to AI irrigation: tech startups powering Latin America’s green revolution. The Guardian, London, 30 jan. 2024. Disponível em: https://www.theguardian.com/global-development/2024/jan/30/low-carbon-milk-to-ai-irrigation-tech-startups-powering-latin-americas-green-revolution. Acesso em: 01 jul. 2025.
WEBB, M. The Impact of Artificial Intelligence on the Labor Market. Stanford, CA: Stanford University, Jan. 2020. [online]. Available at: https://web.stanford.edu/~mww/webb_jmp.pdf. Accessed on: 1 Jul. 2025.
WORLD BANK. East Asia and the Pacific Economic Update: Technology and Jobs in EAP. Washington, DC, oct. 2024. Available at: https://documents1.worldbank.org/curated/en/099101724124028337/pdf/P507415-31cc38ce-ed8a-43a8-9e9c-4d8868dd8c33.pdf. Accessed on: 1 Jul. 2025.
