Reviews of Related Literature

David Fredelique A. Bucoy

This chapter contains ideas, published articles, generalizations or findings, and more.

Those covered in this chapter aid in familiarization with information that is related and similar to

the current study.

Related Literature

In its capacity of “intelligent machine,” AI can carry out typically human tasks, like

pattern recognition, text analysis, and prediction without supervisions. As such, it is akin to other

forms of labor-substituting automation technologies studied by the vast literature recently

surveyed by Restrepo (2023). As discussed in detail there, the “task model of automation”

provides a simple foundation to assess the effect of labor-displacing technological change. The

task model postulates that producing goods and services requires the completion of tasks that can

be assigned to groups of workers, defined by their skills and abilities, or robots and machines.

The crucial takeaway of this framework is that the relative labor demand for groups of workers

whose tasks are taken over by robots falls with automation, and the level of labor demand for

these groups might even fall depending on the productivity gains that come with robot adoption,

and workers’ ability to transition to occupations unaffected by task displacement.

 Acemoglu and Restrepo (2022) derive explicit closed-form log-linear formulas that

summarize the effect of automation on wages of affected worker groups. Aggregate productivity

and the demand for labor in non-automated tasks that workers can easily reallocate positively

impact wages, while the share of tasks displaced by automation directly lower wages. The

overall effect of automation on labor market outcomes thus hinges on whether productivity

improvements and their demand spillovers to other sectors outweigh direct substitution effects.

When applied to the data, this framework explains 50% of the labor share declines across US

industries over the period 1987–2016.

We refer the interested reader to Restrepo (2023) for a full discussion of the automation

literature concerning industrial robots. For the purposes of our review, we can summarize the

empirical literature in this strand as showing that groups experiencing large task displacement

from robot adoption saw decreasing employment and wages, which is consistent with US

evidence that firms deployed these technologies primarily with the aim of substituting human

labor (Acemoglu, Anderson, et al. 2022). Similarly, firm-level studies in various countries

generally found that automation led to a reduction in the labor share and an increase in sales per

worker, although sometimes accompanied by an employment expansion. However, as

highlighted by Restrepo (2023), these effects are hard to interpret and translate to the aggregate

level as adopting firms might just be expanding at the expense of competitors. Accordingly, the

impact of this first wave of automation on economy-wide productivity remain uncertain.

 The above discussion clarifies why the emerging literature on AI has focused on finding a

measure of AI task exposure for different occupations, defined as the share of tasks that could be

carried out by AI in the relevant occupation or job. Under the lenses of the automation literature,

this measure should give a sufficient statistic to predict relative wage changes, as more exposed

workers would see relative wage losses relative to less exposed workers. However, the nature of

AI and specifically gen-AI puts into question whether this technology will act as a complement

or a substitute for workers in related tasks (see, e.g., Autor (2022)). For this reason, the pre-gen-

AI literature on ML has looked to empirical evidence on the relation between AI exposure or

adoption measures and various outcomes, while more recent studies produced experimental

evidence on gen-AI. In what follows, we first present studies on exposure measures and then turn

to the empirical evidence that may be used to interpret these measures, and understand whether

AI might differ from previous waves of automation.

The literature on non-AI automation, focusing primarily on industrial robots, found

strong exposure for blue-collar and less educated workers (Acemoglu and Restrepo 2022), which

has been compounded by increasing substitutability of low-skilled workers with machines (A.

Berg et al. 2024). Under the lenses of potential task exposures, however, AI appears different.
 Correcting for Complementarity. While some of the studies described above explicitly

interpret AI exposure as a measure of potential substitutability, most are agnostic on what

exactly task exposure entails. By contrast, Pizzinelli et al. (2023) construct an indicator of AI

occupational exposure adjusted for the complementarity of the technology with the exposed

occupation (C-AIOE). The resulting measure therefore provides a more direct gauge of the

potential displacement from AI. Occupations’ substitutability and complementarity are assessed

independently from exposure itself based on US O*NET occupation characteristics. 5 The authors

first compute an unadjusted exposure index, reporting that AEs have a higher share of high-

exposure jobs with both high and low levels of complementarity.

When considering demographic characteristics, female and high-skilled workers appear

more exposed. Adjusting for complementarity, the C-AIOE measure shows a lower exposure for

high-skilled occupations relative to its unadjusted counterparts, leading to a more uniform

exposure of AI throughout occupations. Employing this framework in model-based scenarios,

Cazzaniga et al. (2024) find that AI would raise wage inequality for sufficiently high

complementarity, and wealth inequality regardless of complementarity. Similarly, Huang (2024)

classifies tasks that are exposed to AI into groups that allow a better assessment of whether the

technology will substitute or complement human labor. Using this classification, the author

shows that high-skill workers with a high AI exposure would see their employment share in the

economy increase relative to less AI-exposed. This increase would be similar to that experienced
 by progress in IT, which affects predominantly low-skilled workers, but substitutes them

instead of complementing them.

Kogan et al. (2023) provide an estimate of wage and employment effects of AI on

occupational groups that explicitly incorporates the potential of AI to both augment and displace

tasks. The authors calibrate a model that matches empirical evidence on the impact of labor-

saving and labor-augmenting technologies patented since the 1980s. In the model, technology

substitutes routine tasks, augments non-routine tasks, and causes skill obsolescence for

incumbent workers who cannot efficiently use it. Accordingly, each technology has both a labor-

saving and a labor-augmenting exposure. Assuming that workers with the highest labor-saving

exposure will lose 10% of their earnings from AI, the model computes the direct effects–that is,

excluding reallocation across sectors–of current AI technology by occupational group. On

average, workers stand to lose 4% of their earnings. This effect stems from skill obsolescence as,

for almost all occupations, the potential of task augmentation undo the negative task

displacement effects. In terms of skill content, more routine occupations such as office and

admin work, production and transportation stand to lose the most (6–8%), while higher-skilled

occupations like management, engineering, legal, education and social service occupations log

the lowest losses (around 2%).

 Estimates of Aggregate Employment Impacts. Think tank and private sector estimates

provide some additional information on the potential extent of AI impacts on the labor market.

Pew Research (2023) finds that 19% of all US workers are in occupations most exposed to AI. 7

By the same token, Briggs and Kodnani (2023) find an employment exposure of 25% for the US,

24% for the euro area, and 18% worldwide, with effects concentrated in AEs. Ellingrud et al.

(2023) offers somewhat smaller estimates, predicting that 8% of hours currently worked will be

automated by 2030 because of gen-AI, ranging from around 16% of all hours in STEM (science,

technology, engineering, and math), education, and business and legal professional occupations,

to 4% in production work and 3% in agriculture. However, when factoring in other structural

factors, these figures translate into slowing—but still increasing—demand for higher skilled

occupations like STEM, and decreased demand only for selected customer-facing and production

occupations.

Aside from Briggs and Kodnani (2023) and Gmyrek, J. Berg, and Bescond (2023), only

Pizzinelli et al. (2023) offer a glimpse into the potential difference of AI impact between AEs

and EMs. The authors use worker-level microdata from 2 advanced economies (AEs: US and

UK) and 4 emerging markets (EMs: Brazil, Colombia, India, and South Africa). Stronger high-

skilled workers’ exposures also translate into advanced economies being more susceptible to
 employment displacement from AI compared to EM counterparts. However, this

difference is smaller when considering task complementarity.

Economic Feasibility of Task Replacement. In closing this section, it should be noted

that task exposure measures the theoretical possibility that AI may carry out some tasks. This

feature makes exposure a potentially poor measure of the employment effects that we might

expect from the adoption of the technology. Svanberg et al. (2024) explicitly assess the technical

and economic feasibility of automating vision tasks that have a high theoretical AI task exposure.

They find that, given current costs, most US businesses would not automate most tasks with high

exposure. Only 23% of exposed jobs (8% of total jobs) in U.S. non-farm businesses have at leas

one task that is economically attractive to automate. When considering the contribution of

automated tasks to compensation, the authors conclude that only 0.4% of non-farm wages would

be potentially lost to automated computer vision.

Furthermore, in the Philippines, particularly in regions like Surigao del Sur, artificial

intelligence (AI) is beginning to reshape agricultural practices. A study by Maray et al. (2022)

investigates the application of convolutional neural networks (CNNs) for classifying crops,

illustrating how these technologies can help farmers enhance their productivity and forecast

yields more accurately. The findings indicate a noticeable trend among local farmers toward

adopting AI-driven methods to improve their farming techniques. However, the study also

identifies significant hurdles, such as the lack of access to necessary technology and the pressing
need for training programs to help farmers effectively use AI tools. This situation highlights a

broader narrative in the literature on AI in agriculture, which points out that while AI has great

potential to transform the sector, its successful integration requires adequate support and

resources to enable farmers to fully benefit from these innovations.

References

Acemoglu, D., & Restrepo, P. (2022). Tasks, automation, and the rise in U.S. wage inequality.

Econometrica,90(5),1973-2014.

[Link]

rise-us-wage-inequality
Acemoglu, D., Anderson, S., et al. (2022). Artificial intelligence and jobs: Evidence from online

vacancies. American Economic Review, 112(5), 1616-1652. [Link]

id=10.1257/aer.20211396

Restrepo, P. (2023). The economics of automation: Labor substitution and productivity gains.

Annual Review of Economics, 15, 321-348.

[Link]

Berg, A., Chen, K., et al. (2024). Labor market impacts of AI: A comparative study of low-

skilled workers. Journal of Labor Economics, 42(2).

[Link]

Pizzinelli, C., Lee, D., et al. (2023). AI occupational exposure and labor displacement:

Correcting for complementarity. Journal of Economic Perspectives, 37(1), 89-112.

[Link]

Cazzaniga, R., Restrepo, P., et al. (2024). AI and wage inequality: A complementarity-based

analysis. Quarterly Journal of Economics, 139(2), 475-509.

[Link]

Huang, R. (2024). Task substitution or complementation? Analyzing AI’s role in labor markets.

Economic Inquiry, 62(3), 345-367.

[Link]
Kogan, D., Meyer, S., et al. (2023). AI and wage effects: Augmentation vs. displacement.

National Bureau of Economic Research (NBER) Working Paper No. 30556.

[Link]

Maray, D. G., Alcon, R. M., & Caguimbal, J. J. (2022). Convolutional neural network-based

image classification for improved crop production in Surigao del Sur. Journal of Agricultural

Technology,15(3),157-168.

[Link]

Based_Image_Classification_for_Improved_Crop_Production_in_Surigao_del_Sur