My publications
Find here a selection of my recent work on the impact of AI on employment, occupational structures, and labour markets, alongside earlier research on multilateral funding and international development. Each entry includes a brief overview and a link to the full online version.
Navigating the Age of AI, 2026
Implications for Poland’s Economy - What AI Means for Growth, Firms and Jobs
Joint contribution with the World Bank. Poland is the first country where the Bank has applied its new macroeconomic AI framework, built by integrating the ILO's AI Exposure Index (Gmyrek et al., 2025) into the Bank's Computable General Equilibrium model — the same framework used across more than 50 countries. I worked with the World Bank's regional and global modelling teams to adapt the model, translating occupational AI exposure into economy-wide productivity, wage, and employment effects.
The resulting report finds real productivity upside from AI adoption in Poland, with modelled scenarios pointing to meaningful gains in output and wages. But it is equally clear-eyed about the risks: worker displacement concentrated in the most exposed sectors, exposure that falls unevenly across firms and worker groups, and the possibility that productivity gains accrue to some without reaching others. The report's central argument is that this balance isn't decided by the technology itself — it depends on labour market institutions, social protection systems, and social dialogue actively shaping how the transition plays out.
The full methodology behind the integration is being documented in a forthcoming joint World Bank-ILO working paper.
Full report: Navigating the Age of AI: Implications for Poland's Economy, World Bank.
Disruption Without Dividend, 2026
How the digital divide and task differences split GenAI’s global impact
Joint paper with the World Bank, examining how generative AI's labour market effects diverge across regions and income groups. We cover 135 countries and nearly two-thirds of global employment.
While developing economies show lower aggregate automation exposure than advanced economies, digital infrastructure gaps create a stark asymmetry: workers in exposed jobs are usually connected enough to experience displacement, while those positioned to gain from AI-driven task augmentation often lack the access to realize it. The result is that disruption may reach developing countries well before any productivity dividend does.
The paper also shows that standard exposure measures overstate GenAI's impact in developing countries by assuming occupations involve the same tasks everywhere. Drawing on skills survey data from PIAAC, we find that workers in developing countries perform substantially fewer of the non-routine analytical tasks GenAI primarily targets — even within occupations rated as highly exposed. This means exposure estimates need to account for how far a country's job content sits from the technology frontier, not just its occupational mix.
Selected as a background paper for the World Bank's World Development Report 2026 on AI. The paper comes with a full Reproducibility Package available on WB's portal.
Does a General-Purpose Large Language Model Improve Physicians' Clinical Reasoning?
A randomized controlled trial with 249 physicians across Indonesia, Kenya, and the Netherlands
We test whether access to a general-purpose LLM improves clinical reasoning, and what that means for how these tools should be used in healthcare more broadly. LLM access raised performance on standardized clinical vignettes in all three countries, with the largest gains in Kenya (+18%), followed by Indonesia (+10.7%) and the Netherlands (+7.2%).
The results were far from uniform: performance distributions overlapped substantially, and some physicians with LLM access actually performed worse than those without, so access alone didn't guarantee better outcomes. Higher usage tracked with stronger results, and less specialized physicians benefited most, suggesting these tools can help narrow skill gaps rather than just reinforce existing expertise.
The paper is clear that LLMs function as complements to clinical judgment, not substitutes, and flags real risks (automation bias, hallucinations, and context misalignment) that make careful integration and training essential rather than optional.
The paper's policy recommendations go beyond simply granting access: structured integration as a decision-support tool, targeted training, local validation, safeguards against automation bias, investment in infrastructure, continuous monitoring, clear liability frameworks, and inclusive governance to keep deployment equitable and context-appropriate. It also makes the case for social dialogue in shaping how these tools enter clinical practice.
Gen AI, Occupational Segregation and Gender Equality, 2026
We examine how generative AI may reshape gender equality at work, drawing on a global index of GenAI exposure and harmonized microdata from 84 countries.
The core finding: female-dominated occupations, concentrated in clerical, administrative, and business support roles, are almost twice as likely to be exposed to GenAI as male-dominated ones like construction, manufacturing, and trade (29% versus 16%), and face far higher automation risk (16% versus 3%).
Exposure varies sharply by region and income level: 41% of jobs are exposed in high-income countries versus 11% in low-income ones, reflecting differences in occupational structure, digital readiness, and skills. Within that pattern, women are more exposed than men in 88% of countries in the sample.
The higher exposure traces back to entrenched occupational segregation: discriminatory norms and biases in recruitment, promotion, and workplace practices that shape who ends up in which jobs in the first place.
Generative AI and Jobs, 2025
A refined index of occupational exposure
We surveyed over 1,600 workers in Poland about how AI could affect specific tasks in their jobs, and compared their views with insights from international experts. Using this data, we trained an AI model to estimate exposure levels across all occupations at the global level.
The results show that clerical and highly digital jobs are most exposed, with about one in four workers globally in roles likely to be affected. Women and workers in high-income countries are especially exposed. While few jobs are fully automatable, many will change — and understanding these shifts is essential for shaping fair and informed policy responses.
A new chapter for the ILO’s textual assets
Applying Generative AI to Labour Force Survey questionnaires
This work shows how AI can help transform existing textual resources into valuable digital assets for research, policy, and future innovation. We used AI tools to extract, digitize, and organize thousands of Labour Force Survey questions from complex forms, combining machine vision with human checks for accuracy. The result is a searchable database powered by AI, allowing users to quickly find and understand survey content across countries and years. We also built an interactive search app for ILO users and published the full process — including Python code — on GitHub.
Buffer or Bottleneck?
Generative AI and the Digital Divide in Latin America
Joint paper with the World Bank. We investigate how generative AI could affect jobs across Latin America. Using harmonized labour force and household survey data, we estimate GenAI exposure across countries and sectors, adjusting for differences in digital access and technology adoption. Our findings show that the digital divide remains a major barrier: nearly half of the jobs that could benefit from GenAI remain excluded due to limited access to digital tools in the workplace.
The paper comes with a full Reproducibility Package available on World Bank's portal.
A Technological Construction of Society
Comparing GPT-4 and Human Respondents for Occupational Evaluation in the UK
In this paper, we compare how GPT-4 and humans evaluate occupations in the UK. Using large survey data covering 580 occupations and two key metrics — prestige and social value — we find strong overall alignment between GPT-4 and human assessments. However, the model tends to misjudge certain roles, especially emerging digital jobs and stigmatized or illicit occupations. We also show that GPT-4 consistently imitates majority opinions and struggles to reflect the views of minority groups. These findings highlight both the potential and limitations of using large language models in sociological and labour market research, with important implications for policy and inclusion in the future of work.
Generative AI and the media and culture industry
We explore how generative AI is reshaping jobs in the media and culture industry. GenAI is transforming creative processes in journalism, music, film, and beyond — not only how content is produced, but also who holds creative control. Some roles, like writing and translation, face high exposure to automation, while performance-based jobs remain largely human-centred. We highlight the growing demand for skills in AI tool management, ethics, and digital literacy, alongside traditional creative strengths. We also examine policy responses to ensure fair compensation, protect creative agency, and support workers through this transition.
Who cares about workers’ rights?
The effects of violations of trade unions’ rights on donors’ funding decisions in the ILO
I examine how respect for trade union rights influences donors’ aid decisions to the ILO. Using data on multi-bilateral contributions and labour rights indicators, I find no systematic link between the level of rights violations in a country and the volume of aid it receives. Still, many ILO-funded programmes in countries with the most severe violations focus on strengthening workers’ organizations. The paper raises important questions about how aid should be allocated in these contexts, and whether it should be more closely tied to international labour standards and oversight mechanisms.
Trade interests and UN funding
Commercial earmarking of multi-bi aid
A full-length book published as part of the Routledge series on international organizations. I examine how donor countries use earmarked aid within the UN system, and how commercial interests can influence funding decisions. Focusing on country-specific earmarking, the book traces how these patterns developed during the MDG era and explores their implications for the current SDG framework. While commercial motives are often viewed critically, I argue that their impact on multilateral development cooperation depends on the availability of flexible funding and the institutional autonomy of UN agencies. The book offers a structured and accessible introduction to the political economy of multi-bi aid and its role in global governance — particularly relevant in an era of renewed focus on national self-interest, trade-driven diplomacy, and declining development aid.
