Occupational, Industry, and Geographic Exposure to Artificial Intelligence: A Novel Dataset and Its Potential Uses

By Edward Felten, Manav Raj, and Robert Seamans

A large and growing body of research has generated both excitement about artificial intelligence’s (AI) potential to transform the nature of work and concerns about mass unemployment. However, despite all the press and attention paid to AI, there has been little systematic collection of evidence, partly because the field is so new, and partly because appropriate tools to measure its impact have yet to be developed.

A new paper addresses this shortcoming and introduces a novel dataset that measures the exposure to AI applications across occupations, industries, and geographic regions. The dataset has been made available so that other researchers can try to answer some of the many open questions about the effect of AI on workers, firms, and markets.

The dataset looks only at “narrow” AI or machine learning (i.e., computer software that relies on highly sophisticated algorithmic techniques to find patterns in data and make predictions about the future) and does not address robotics or automation more broadly. The dataset is unique in that it measures the degree to which a particular occupation is exposed to AI, without regard to whether AI substitutes for or complements human work. It also includes a new method for linking AI applications to human abilities. 

The dataset links ten common AI applications with fifty-two distinct occupational abilities used by O*NET and uses the relative prevalence of these abilities within an occupation to generate an occupational measure of AI exposure. The AI occupational exposure (AIOE) measures an occupation’s relative exposure to AI, and captures how the most prevalent applications of AI are related to occupations based on their ability composition as defined in O*NET. 

With the AIOE as a baseline, the data can be aggregated across an industry, to determine the industry-wide AI exposure (AI Industry Exposure, or AIIE), or across geographic areas, using county-level data to determine the AI Geographic Exposure (AIGE). These measures are flexible and can be updated to reflect new titles and occupation descriptions in O*NET, or the relative prevalence of occupations across an industry or geographic region.

A few examples help validate the data on an intuitive level. The occupations with the highest exposure to AI — again, without regard to whether the AI complements or substitutes — include genetic counselors, financial examiners, and actuaries, all occupations that typically require advanced degrees. Those occupations with the lowest exposure include non-office jobs with a substantial physical component, including dancers, painters, and fitness instructors. The industries with the highest AI exposure include financial services, accounting, insurance, and legal services. Those with the least exposure to AI include industries with substantial manual labor components, including crop production, construction, and building services. All of these results are consistent with prior research and common sense.

Comparison of some jobs with disparate AI exposure measures further validates the data. For example, surgeons and meat slaughterers both require deft physical manipulation of human or animal tissue and require a high level of skill in manual dexterity, finger dexterity, and arm-hand steadiness. Surgeons have a much higher AI exposure, however, because of the importance of various cognitive abilities, such as problem solving, deductive and inductive reasoning, and information ordering, that are not as important for meat slaughterers. Thus, the varying AI exposure scores for surgeons and meat slaughterers are borne out by the intuitive understanding of the cognitive abilities required in each. 

Similarly, mathematical technicians and accountants both require a high degree of cognitive skill, but accountants have a much higher AIOE — the 99th percentile — than mathematical technicians, which score in the 68th percentile. The difference stems from the presence of sensory versus physical or psychomotor abilities. Cognitive abilities are more exposed to AI relative to all other abilities; however, differences remain across noncognitive abilities. In particular, sensory abilities, defined as abilities that influence visual, auditory, and speech perception, are more exposed to AI technologies than physical or psychomotor abilities. This is perhaps not surprising given that AI technologies are particularly known to be well-suited for tasks involving classification, categorization, and pattern recognition, and perception may be keenly involved in such tasks. Of the noncognitive abilities required for accountants and auditors, 90.6% are sensory abilities, while for mathematical technicians, only 52.6% are sensory abilities. The higher relative importance of sensory abilities (vs. physical or psychomotor abilities) for accountants and auditors results in the higher exposure to AI relative to mathematical technicians, despite a similar level of cognitive abilities.

These measures are just a start for potential uses of this new dataset.The data may be used, for example, for studying competitive and corporate strategies around AI, organizational design, human resources management, industrial organization, geography of innovation and spillovers, regional and business dynamism, and entrepreneurship. This dataset could also be modified to explore AI exposure at the firm level, using available microdata from other sources, and it could shed light on the relationship between AI, human capital, and labor productivity. All told, this novel dataset begins to fill a gap in the analysis of the effects of AI, and will hopefully lead to further understanding as AI applications continue to develop. 

Strategic Management Journal 2021: 1-23.

Dataset on GitHub.