Which Jobs Will AI Take Over?

Summary

• A new NBER working paper reveals a recent surge in job listings requiring AI-specific skills, based on Burning Glass job vacancy data.
• Firms are hiring not only more AI-related workers but also disproportionately fewer non-AI workers.
• Finance, business services and information-related sectors are most vulnerable to AI disruption.
• Although there are net job losses from AI adoption, they have yet to impact economy-wide employment trends.

Introduction

A new NBER working paper explores how AI has impacted US jobs. Using Burning Glass data from 2010-2018, authors Daren Acemoglu et al. document a rapid growth in AI-related vacancies, particularly in ‘AI exposed’ establishments. They find the following:

1. AI exposure is associated with both a significant decline in some skills previously demanded in vacancies and the emergence of new ones.
2. Task substitution partly drives the recent AI surge, whereby AI automates a subset of tasks formerly performed by labour.
3. The researchers detect no relationship between AI exposure and overall employment or wages at the industry or occupation level.

The paper shows that job losses from AI are larger than job gains in firms that have so far adopted AI. Nevertheless, the aggregate impact is still too small to have had first-order impacts on US employment patterns.

Data

Burning Glass collects data from roughly 40,000 company websites and online job boards. The vacancy data includes occupation, industry and region information, firm identifiers, and detailed information on skills required by vacancies. This is all garnered from the text of the job posting. The Burning Glass data closely tracks overall vacancies in the BLS Job Openings and Labour Turnover Survey (JOLTS).

The authors collect data on vacancies by firm name, location, and industry code (each firm is classified according to the industry in which it posts the most vacancies over the sample period). Within this sub sample, they look at vacancy details relating to skills and occupations. They create two measures of AI vacancies: a narrow one that includes skills relating to AI tasks, and a broad one that uses ‘skill clusters’ relating to AI tasks.^[1] Both measures highlight a rapid uptick in AI vacancies since 2016 (Chart 1).

Source: Paper, page 31

AI Exposure

The authors use three measures of ‘AI exposure’. Firms are classified as AI exposed when their workers engage in tasks that are compatible with current AI capabilities. These are the measures:

1. Felten et al. (2019) – based on data from the Electronic Frontier Foundation. The data identifies a set of nine application areas in which AI has made progress since 2010.
2. Brynjolfsson et al. (2018) – a 23-item rubric for systematically scoring the suitability of any given job task for machine learning/AI (named the Suitability for Machine Learning (SML) measure).
3. Webb (2020) – identifies occupations that are exposed to AI based on whether the occupation’s tasks overlap with capabilities described in AI patents. Occupations with a larger fraction of overlapping tasks are classified as more exposed.

Greater emphasis is placed on measures one and three because they are strongly predictive of AI adoption. In other words, the more tasks within a company that can be done through AI, the more likely the company is to post AI vacancies. Specifically, a one standard deviation increase in AI exposure (which corresponds to the difference between the finance and the mining and oil extraction sector) is associated with approximately a 16% increase in AI vacancies.

Chart 2 shows the degree of AI exposure by broad occupation and one-digit industry. The Felten et al. measure, for example, is particularly high for managers, professionals and office and administrative staff but is very low for service, production and construction workers. This captures the fact that these occupations involve various manual tasks that algorithms cannot perform.

Overall, the most vulnerable sectors to AI adoption are finance, business services, information-related sectors and manufacturing. And in terms of occupations, managers, professionals and administration roles are most at risk.

Source: Paper, page 32

AI Is Changing the Types of Skills Businesses Need

It seems reasonable to assume that firms with task structures suitable for AI will tend to change the types of skills they demand. On this front, the authors find that the deployment of AI technologies goes hand-in-hand with significant skill redundancies:

• A one standard deviation increase in AI exposure is associated with a 0.83 absolute decline in listings that require the skills previously demanded.
• A one standard deviation increase in AI exposure is associated with a 0.95 absolute increase in listings that require skills not previously demanded.

This, the paper suggests, confirms two things. First, that AI exposure is associated with significant changes in the skill sets posted in vacancies. Second, that it is enabling firms to replace some of the tasks that workers previously performed. This makes certain skills redundant, while simultaneously generating demand for new skills (‘skill churn’).

Firms More Likely to Implement AI Are Less Likely to Hire in Non-AI Roles

One question that remains is whether the churn created by AI complements labour or substitutes for it. On this front, the authors consider two time periods, 2010-2014 and 2014-2018. Within the first five years, AI adoption had not yet taken off and, reassuringly, there exists no significant relationship between AI exposure and non-AI hiring.

Between 2014 and 2018, however, the relationship shows that a one standard deviation increase in AI exposure is associated with an approximately 12% decline in overall non-AI vacancies. This suggests that the surge in AI vacancies is likely to have had significant and economically meaningful negative effects on hiring.

If jobs are being displaced from AI-exposed firms, then perhaps less exposed firms are absorbing them. In other words, there may also be changes in industry-level organisation that potentially offset or augment the establishment-level consequences. Indeed, a rise in industry-level AI exposure has no overall effect on industry employment. This suggests either that it is too soon to see the impact of AI activity on industry reorganisation or that jobs lost within one part of an industry are absorbed by other similar firms.

Furthermore, the authors find that changes in employment and wages across occupations with varying levels of AI exposure are insignificant. In other words, they detect no relationship between AI exposure and overall employment or wages at the industry or occupation level.

Bottom Line

Many of the results from this paper seem unsurprising – AI has so far had no aggregate impact on employment or wages, and the number of AI vacancies is rising. Nevertheless, two key results stand out: (i) the increase in listings requiring AI-related skills has coincided with a decrease in listings requiring non-AI-related skills, and (ii) firms more likely to implement AI are less likely to hire in non-AI roles.

These outcomes show that AI is having real effects. There is a significant surge in vacancies for AI workers in the US, and simultaneously there are non-trivial reductions in non-AI hiring in more AI-exposed establishments. Although this appears to be happening only within specific industries, as adoption of AI increases, the substitutability of worker for AI could lead to a hollowing out of the workforce.

Citation

Acemoglu D., Autor D., Hazell J. and Restrepo P. (2020), AI and Jobs: Evidence from Online Vacancies, NBER Working Paper 28257, https://www.nber.org/papers/w28257

1. The broader measure is likely to include skills that are separate from core AI activities, such as various IT functions. ↑