AI and Employment / In the Era of Generative AI, Whose Jobs Will AI Reach? / Reflections on Reading the EU JRC Empirical Paper

I. The Resurgence of 'Automation Anxiety' and the Limitations of Previous Research

The estimates published by Frey and Osborne of the University of Oxford in 2013 concluded that 47% of U.S. employment was in a state of 'high risk of automation,' and this figure has been repeatedly cited in media and policy discussions ever since [Note 1]. Subsequently, the OECD re-analyzed the data at the task level, providing revised estimates suggesting that even with similar methods, the risk of automation remained at around 9% on average [Note 2]. The conclusion that 'high risk is primarily in low-skilled occupations' had formed a general consensus across these two research lineages.

However, following the release of ChatGPT in November 2022, Generative AI (GenAI) based on Large Language Models (LLMs) rapidly expanded its capabilities in text generation, summarization, and logical reasoning. GenAI is now directly encroaching upon parts of 'creative intelligence' and 'social intelligence,' which the Frey-Osborne study had positioned as 'bottlenecks to automation.' Whether the exposure indices constructed before GenAI remain valid must be re-examined.

A working paper published by the European Commission's Joint Research Centre (JRC) in 2026 is a systematic attempt at such a re-examination [Note 3]. It connects 352 AI benchmarks to 127 ISCO-3 occupations via 14 cognitive abilities and 108 job tasks, constructing a dynamic AI exposure index spanning from 2008 to 2024. Below, I will organize its methodological framework and key findings, and then consider what can be gleaned from them.

II. What is the Three-Layer Mapping?

1. The Significance of Measuring AI Progress by 'Number of Papers'

How to quantify the progress of AI research is not self-evident. Even if one compares accuracy improvements on individual benchmarks, cross-sectional evaluation is difficult because the difficulty and nature of the benchmarks differ. This paper adopts an approach that uses the number of academic papers published in relation to specific benchmarks as a proxy indicator for research intensity. The design is based on the premise that the more a research community concentrates its efforts on a field, the more likely it is that progress will occur in the near future; it is accurate to understand this not as a direct measurement of progress, but as a measure of where the 'gravity' of AI research is heading.

By logarithmically transforming the number of papers and constructing them as cumulative time-series values, the index is not swayed by single-year fluctuations, allowing us to capture how much cumulative research AI has attracted for each cognitive ability over which period.

2. The Role of the Intermediate Layer of Cognitive Abilities

The core of the methodology lies in not connecting AI benchmarks and job tasks directly, but by interposing an intermediate layer of 14 cognitive abilities. This framework is set as general abilities common to both humans and AI, and is broadly categorized into three groups: 'Idea-related,' 'People-related,' and 'Object-related' [Note 3]. By comparing AI benchmarks and job tasks along the same axis of cognitive abilities, we can structurally visualize the pathways through which technological progress reaches occupations.

The correspondence between the 108 tasks and 14 cognitive abilities was determined using a Delphi method combining two human experts and four LLM agents. This paper also conducts a detailed verification of this process, experimentally showing that results differ significantly depending on whether AI functions as a complement to humans or as a substitute. The description that the substitution experiment ended in failure and only the complement model reached a consensus can be evaluated as methodological integrity, while simultaneously resonating with the policy implications discussed later.

III. The Structure of Findings

1. Concentration on 'Idea-related' Cognitive Abilities and Their Transversality

Recent progress in AI research is strongly skewed toward specific cognitive abilities. The areas where exposure is accelerating the most are attention and search, comprehension and expression, conceptualization/learning/abstraction, and quantitative/logical reasoning [Note 3], which correspond to the domains that LLMs and vision-language models have directly targeted for design. In the AI exposure score as of 2024, these abilities show the highest medians and the largest variances.

Conversely, abilities related to social cognition and physical interaction—such as emotion regulation, mental modeling, metacognition, sensorimotor interaction, and navigation—have remained low in both exposure levels and growth rates from 2012 to 2024.

What is important here is that the 'idea-related' cognitive abilities where exposure is most concentrated are not unique to professional occupations. Cognitive functions such as understanding instructions and basic information retrieval are used daily even in occupations considered to be simple labor. The areas where technological progress has concentrated happened to overlap with abilities that are distributed transversally across the entire labor market. This is the structural reason that brought about the ripple effect of exposure.

2. A Shift Toward Ripple Effects Across All Occupational Hierarchies

Since 2015, AI exposure has risen sharply in all occupational categories. It remains unchanged that managers, professionals, and technicians show the highest levels of exposure. However, exposure scores are also continuously rising in clerical, service and sales, and even elementary occupations, reaching job groups that research prior to GenAI had classified as having 'weak exposure' [Note 3].

What is particularly important as a numerical fact is that the average AI exposure level of low-skilled jobs as of 2024 exceeds the level of high-skilled jobs as of 2018. While the hierarchy between job types remains, the phenomenon of exposure itself has come to traverse all layers of the labor market. The composition of 'low-skilled jobs are at high risk' depicted by the Frey-Osborne study is no longer an accurate map in the GenAI era.

3. The Correlation Between Exposure and Income, and Reservations in Its Interpretation

As of 2024, the correlation coefficient between the AI exposure score and the average income decile of occupations is 0.85, confirming a strong positive correlation [Note 3]. Higher-income occupations have higher AI exposure. This is the exact opposite of the composition shown by the Frey-Osborne study.

However, I must emphasize again that what this paper measures is the technical degree of alignment between task content and the direction of AI research, not a prediction of whether actual substitution or complementarity will occur. The authors also explicitly reserve this point, acknowledging that the same exposure can turn into either substitution or complementarity for workers in different categories.

IV. What Distinguishes Substitution from Complementarity

The fact that an exposure score is high does not mean that an occupation will be immediately replaced by automation. For professionals, AI can become a means to complement high-level intellectual work and improve productivity. On the other hand, if similar exposure occurs in clerical or low-skilled jobs, concerns arise that it will lead directly to the substitution of labor through task automation. Even with the same technical exposure, the results produced can be reversed depending on how labor is organized in each company and the distribution of workers' bargaining power.

It is precisely in this undecided area that there is room for law and social institutions to intervene. As a policy implication, the authors point out the importance of inducing AI implementation to function as a complement through the strengthening of collective bargaining structures, in addition to education and training policies [Note 3]. Whether the benefits brought by technological progress reach all strata is a matter of institutional design between labor and management.

V. In Conclusion

The most important shift in perception presented by this paper is that AI exposure is transitioning from a 'problem of high-level occupations' to a 'problem of the entire labor market.' Some of the capabilities that pre-GenAI research considered protected areas are rapidly becoming the target of AI research, and many of the cognitive tasks performed daily in low-skilled occupations are entering the sphere of exposure.

The expression 'AI as a tide can lift all boats' appears near the conclusion of this paper, but the authors simultaneously show with figures that the height of the waves still maintains inequality. The direction in which the technical fact of exposure turns depends on the design of implementation in the workplace and the distribution of bargaining power. The question of how to reconstruct the value of labor and the distribution of social wealth must be repositioned as a challenge for law and policy, not as a matter of technology. Please check the original source for details. I hope this provides some food for thought.

(Magazine) "Reflections on AI and Law"

*Please refer to the following for the table of contents

note General Terms and Conditions Article 3, Paragraph 2, First Sentence
3.2 The copyright of digital content produced by the creator belongs to the creator.
P.S. This is shared for academic discussion only.