Generative AI and the labor market: A case for techno-optimism

Gen AI can enhance labor demand and productivity as well as ease wage inequality, yet it remains to be seen if everyone can reap its many benefits.

Ira Kalish

United States

Michael Wolf

United States

Technological advancement and innovation have a long history of creating anxiety among workers. In the early 1800s, the so-called Luddites smashed knitting machines that threatened textile workers in England.1 In 1930, John Maynard Keynes posited that technological innovation could at least temporarily result in widespread unemployment.2 Today is no different. Plenty of ink has been spilled fretting about robots replacing human labor. Although it is easy to see the displacement effect that technology can have on labor, innovation can also raise demand for labor in other sectors, making the net effect a positive one.

The historical record is mixed when it comes to innovation and its impact on the labor market. Sometimes, innovation can have a positive effect on labor demand, while at other times, it can be detrimental. So, how can we know what type of effect generative artificial intelligence will have? And what does it mean for the rest of the economy?

A brief review of the literature on the economics of innovation is useful in understanding the interaction between innovation and the wider economy. A wide assessment of innovation in the United States shows, that for four decades following World War II, innovation strengthened labor demand, while weakening it over the four decades following that time period.3 This does not mean labor demand has fallen over the latter period, but that it is weaker than the counterfactual and previous time period. After all, the employment-to-population ratio in the United States peaked in 2000, well into this slower labor demand period (figure 1).

Stronger growth in the earlier period is often associated with the five “Great Inventions of the Second Industrial Revolution,” which include electricity (for example, the light bulb and electric motor), the internal combustion engine, sanitation (for example, running water and indoor plumbing), chemicals (e.g., natural gas, plastics, and pharmaceuticals), and telecommunications (e.g., the telephone and radio). Some researchers show that these inventions had such a profound and anomalous effect on the economy and living standards that their strong economic effects are unlikely to be repeated in the future.4

The type of innovations that occurred over the 40 years following World War II likely explains some of the associated positive economic outcomes. Most of these inventions would be considered product innovations, which are creations of a new or vastly improved good or service rather than a process innovation, which is focused on how a good or service is created. For example, the invention of the automobile would be a product innovation, whereas the invention of industrial robots that assemble automobiles would be a process innovation. Research shows that product innovations are more likely than process innovations to improve productivity and therefore boost economic output.5

Researchers focusing on the labor market invoke a similar argument.6 They believe technological innovation focused on augmenting labor or performing tasks that humans do not or cannot perform will lead to stronger growth in aggregate labor demand. Such innovations are more closely aligned to product innovations. Conversely, labor-automating technologies that focus on performing existing human tasks, such as aspects of AI, are more similar to process innovations and are more likely to diminish aggregate labor demand.

Making distinctions between product and process innovations, and labor-augmenting and -automating innovations can be murky. For example, the invention of the automobile can be considered a product innovation in that it was sold to customers and was significantly different from anything that came before it. However, it can also be a process innovation when it is used to transport goods. Similarly, an innovation such as a word processor can be considered labor-automating to a legal secretary but labor-augmenting to a lawyer.

The scale of an innovation is perhaps more important than whether it is a product or labor-augmenting innovation. Research focusing on innovations among French manufacturers shows that it is “radical innovation,” rather than “incremental innovation,” that has a positive effect on overall productivity growth.7 Similarly, two leading researchers on the topic note that it is not the “brilliant” automation technologies that threaten employment and wages, but “so-so technologies” that generate small productivity improvements.8 A “brilliant” technology would include the invention of refrigeration, which drastically reduced food spoilage and improved productivity in agriculture and food processing. A self-service kiosk that shifts work from the cashier to the customer without improving quality could be considered a “so-so” innovation.

From a theoretical standpoint, we can decompose innovation’s effect on the labor market into three distinct effects: the displacement effect, the reinstatement effect, and the productivity effect.9 The displacement effect reduces employment as innovation automates tasks and therefore reduces the demand for labor. The reinstatement effect improves labor demand as innovation creates new tasks that humans will need to perform. For example, more data and computer scientists could be employed to produce and maintain an automating technology. Finally, the productivity effect increases demand for labor in unaffected industries as more productive economic activity raises incomes. For example, when refrigeration reduced spoilage, it brought down costs typically shared between businesses and consumers. Those cost savings can then be deployed elsewhere in the economy, driving up labor demand in those other industries.

Even if labor-augmenting and product innovations are preferable to labor-automating and process innovations, respectively, many experts in the field of innovation economics argue that the magnitude of the productivity effect is what will ultimately determine if an innovation increases or decreases aggregate labor demand. The mechanism for this productivity effect is important. Larger cost savings in the innovation firm or industry will yield higher productivity growth. This means that the productivity effect is strongest when wages are high, and labor is scarce in the innovating firm or industry.10

Where does generative AI fit into this?

Numerous innovations are being made every day, but in this article, we focus exclusively on generative AI, which uses foundation models to create new content in the form of text, code, voice, images, etc.11 This technology is still in its infancy, and advancements that can be built based on this technology remain unknown. Even so, based on our understanding of generative AI as it currently exists and what some researchers predict could happen with this technology in the near term, we can assess how it might affect labor markets. For a deeper understanding of generative AI capabilities and business applications, explore the Deloitte AI Institute.

First, understanding why innovation has had a more limited productivity effect over the last 40 years acts as a useful starting point. At least part of the reason why post-1980 innovation had a weaker productivity effect was likely due to the types of workers that were displaced—automation over this period largely focused on low- and middle-wage workers. Word processors and spreadsheets displaced relatively low-paid clerical workers, such as file clerks. Meanwhile, machinery and industrial robots displaced middle-wage factory workers.12 Because the cost savings were relatively low under these circumstances, the proportional productivity effect was also smaller.

This is unlikely to be the case for generative AI, however. Research that matches generative AI skills with those of workers finds that higher-wage workers are the most at risk of losing their jobs to this technology.13 Although the exact occupations that are at risk differ across research methods, there is widespread agreement that the share of tasks that could be done by generative AI rises with income.14 Some of the occupations that have been deemed most at risk of automation from generative AI include postsecondary educators, mathematicians, and survey researchers. The industries with the greatest exposure often include legal, financial, and professional services.15 Most of these occupations and industries involve high wages, suggesting that the cost savings of this technology could be substantial. Assuming the costs saved are indeed proportional to the productivity effect, then we should see stronger—rather than weaker—aggregate labor demand.

We know from the previous section that process innovations have a lower likelihood than product innovations to increase demand for labor. Although generative AI could be considered a new product, it is likely going to be used as a process innovation across most use cases. This alone reduces the probability of generative AI improving labor demand. However, the mixed outcomes of process innovations likely reflect the distinction between “brilliant” technologies and “so-so” technologies. It is the former of these technologies that have the largest productivity effects and boost labor demand. 

Although there is no standard definition of what technologies qualify as “brilliant” or “so-so,” generative AI likely counts as the former. For one, the technology has wide-reaching effects as it is considered a general-purpose technology.16 This means that generative AI has applications across multiple industries and can perform a variety of tasks. For example, generative AI has proved to be adept at writing code for different types of software,17 training telemarketers,18 providing research support,19 and detecting fraud.20 General purpose technologies do not guarantee a strong productivity effect. After all, they could have only a small positive or even neutral effect on productivity growth even if they are deployed widely. But the wide array of tasks it can perform and occupations and industries it will affect, bode well for a strong productivity effect.

There are at least two confounding factors that could support or hinder the expected increase in labor demand from generative AI. The first is demographics. Scarcity of labor not only bids up wages, increasing the cost savings from innovation, but also encourages more investment in automation. Countries and firms most reliant on middle-aged workers that were dwindling in number have historically experienced greater automation to offset the demographic decline.21 Given that most developed countries are seeing weaker or even negative working-age population growth,22 widespread adoption of generative AI becomes likelier. Such widespread adoption then raises the probability of high cost savings, improved quality of output, and therefore, a stronger productivity effect.

The second factor comprises institutions. Most studies covering how innovation affects labor markets focus on the US economy due to availability of data. However, the experience in other countries might look quite different. In countries with stronger worker protection laws and higher rates of union membership, the displacement effect of generative AI is likely to be smaller, at least in the near term. For example, in the United States, the adoption of industrial robots was associated with lower labor demand in the industry and weaker labor demand in the local labor market. However, in Germany, where worker protections are stronger, the adoption of industrial robots had no discernible effect on local labor demand in the decade between 1994 and 2014.23 German manufacturing workers were largely able to stay at their current employer but switched roles internally. The growth of new workers in the affected industry slows down, ultimately offset by gains in the business service sector. More recent union negotiations, such as the writer’s strike in the United States, have focused heavily on protecting workers from AI-related disruption.24

In geographies or industries with stronger protections in place, the displacement effect of generative AI may prove to be more muted or even nonexistent. Under these circumstances, stronger labor demand is more likely as the negative labor market effects of generative AI are diminished. This also means that the productivity effect could be smaller in aggregate since cost savings will see a drop. However, labor protections should allow for a more orderly shift in occupational tasks and smooth out demand. It also shifts firms’ attention away from task automation and toward using generative AI to make productivity gains and improve quality of output—both of which should have positive effects on labor demand.

Pressuring inequality

Assuming the widespread adoption of generative AI creates a stronger productivity effect than its displacement effect, labor demand should increase and unemployment rate should fall. It should also increase total productivity growth, which will raise real per capita incomes, real GDP, and consumer spending. Stronger productivity growth is also associated with lower inflation as capacity constraints diminish. This would help reverse the slow productivity growth seen in developed economies over the last 10 to 15 years.

Like other technological innovations before it, generative AI could have an impact on inequality. Given that the most at-risk occupations and industries are those at the higher end of the income distribution, displacement of these workers should reduce inequality. One research paper shows that inequality between those in the 90th and 10th income percentiles should fall amid adoption of generative AI. However, that same paper shows that there will be little negative effect on those in the top 1% of income distribution.25 Of course, reducing employment of higher-wage workers will reduce inequality. However, replacing high-earning skills with generative AI could expand demand for lower-paid skills, bidding up demand and wages for lower-paid workers and putting additional downward pressure on wages for high skilled workers.

Even if these economic outcomes come to pass, it does not guarantee that everyone will be better off. Regions with high concentrations of affected workers may otherwise be worse off. Returning to the experience of industrial robot adoption in the United States is perhaps a cautionary tale. Those robots reduced labor demand not just in the affected industry or firm but in the entire local labor market. Indeed, numerous US regions with high concentrations of manufacturing jobs that were subsequently automated or offshored faced a litany of economic and social problems thereafter.26 Should generative AI have an outsized effect on workers in a particular industry, the regions with large concentrations of that industry could struggle as a result. Financial services are frequently cited as among the most exposed industries, which could put the economies of global financial centers at risk.

Policy could also change in response to this technology. Critics of labor-displacing technology have suggested that a change in the tax code could encourage more hiring relative to technological investment, thereby impeding negative effects on the labor market.27 Indeed, research shows that there is unfavorable tax treatment of workers relative to investment in the United States.28 If generative AI does not end up having the strong productivity effect we expect, it could increase pressure to raise tax rates on investment or reduce them for labor.

There are still many variables when it comes to generative AI and how it will change the way we do business. Based on what we know so far, it has a high probability of improving labor demand, reversing a four-decade trend of technological innovation weighing on labor. After all, there should be considerable cost savings associated with the technology that will raise the important productivity effect. Stronger labor demand and productivity will result in favorable economic outcomes, including a potential reduction in inequality. This relatively optimistic outlook, however, does not mean everyone will be better off, as some workers and geographies could be overly exposed to the negative effects of the technology.

By

Ira Kalish

United States

Michael Wolf

United States

Endnotes

  1. Kyle Chayka, “Rethinking the Luddites in the age of AI,” New Yorker, September 26, 2023.

    View in Article
  2. Luciano Floridi, “Technological unemployment, leisure occupation, and the human project,” Philosophy & Technology 27 (2014): pp. 143–150. 

    View in Article
  3. David Autor, Caroline Chin, Anna Salomons, and Bryan Seegmiller, New frontiers: The origins and content of new work, 1940–2018, MIT, August 14, 2022. 

    View in Article
  4. Robert Gordon, Productivity and growth over the years at BPEA, conference draft, Brookings Papers on Economic Activity, March 25, 2021. 

    View in Article
  5. Bronwyn H. Hall, Innovation and productivity, National Bureau of Economic Research, accessed December 5, 2023.

    View in Article
  6. Daron Acemoğlu and Simon Johnson, Rebalancing AI, International Monetary Fund, December 2023.

    View in Article
  7. Ibid.

    View in Article
  8. Daron Acemoğlu and Pascual Restrepo, Automation and new tasks: How technology displaces and reinstates labor, IZA Institute of Labor Economics, accessed December 5, 2023.

    View in Article
  9. Andrew Green, OECD employment outlook 2023: Artificial intelligence and the labour market, OECD, accessed December 5, 2023.

    View in Article
  10. Ibid.

    View in Article
  11. Deloitte AI Institute, Generative AI is all the rage, Deloitte, accessed December 5, 2023.

    View in Article
  12. Daron Acemoğlu and Pascual Restrepo, Tasks, automation, and the rise in US wage inequality, MIT, February 15, 2022. 

    View in Article
  13. Tyna Eloundou, Sam Manning, Pamela Mishkin, and Daniel Rock, GPTs are GPTs: An early look at the labor market impact potential of large language models, Arxiv.org, March 23, 2023.

    View in Article
  14. Ibid.

    View in Article
  15. Ed Felten, Manav Raj, and Robert Seamans, How will language modelers like ChatGPT affect occupations and industries?, Arxiv.org, March 18, 2023.

    View in Article
  16. James Pethokoukis, “The economic promise of ChatGPT and GenAI as a general purpose technology,” AEI, April 24, 2023. 

    View in Article
  17. Deloitte AI Institute, The generative AI dossier: A selection of high-impact use cases across six major industries, Deloitte, accessed December 5, 2023. 

    View in Article
  18. Ajay Agrawal, Joshua S. Gans, and Avi Goldfarb, The Turing transformation: Artificial intelligence, intelligence augmentation, and skill premiums, Brookings, June 12, 2023. 

    View in Article
  19. Anton Korinek, “Language models and cognitive automation for economic research,” National Bureau of Economic Research, February 2023.

    View in Article
  20. Padma Chukka, “Generative AI: The missing piece in financial services industry?,” Finextra, April 21, 2023.

    View in Article
  21. Daron Acemoğlu and Pascual Restrepo, Demographics and automation, MIT, January 2021. 

    View in Article
  22. Barry P. Bosworth and Gary Burtless, “Budget crunch: Population aging in rich countries,” Brookings, June 1, 1997. 

    View in Article
  23. Wolfgang Dauth, Sebastian Findeisen, Jens Suedekum, and Nicole Woessner, Adjusting to robots: Worker-level evidence, Federal Reserve Bank of Minneapolis, August 21, 2018.

    View in Article
  24. Rebecca Klar, “How Hollywood writers set a new standard for AI protections,” The Hill, accessed December 5, 2023.

    View in Article
  25. Michael Webb, The impact of artificial intelligence on the labor market, Stanford University, January 2020. 

    View in Article
  26. Carnegie Endowment for International Peace, How trade did and did not account for manufacturing job losses, December 10, 2018.

    View in Article
  27. Diego Areas Munhoz and Samantha Handler, “AI-fueled job displacement anxiety triggers tax code scrutiny,” Bloomberg Law, June 15, 2023. 

    View in Article
  28. Christopher Ingraham, “For the first time, workers are paying a higher tax rate than investors and owners,” Washington Post, October 16, 2019.

    View in Article

Acknowledgments

The authors would like to thank Sue Cantrell, Steven Hatfield, and John O’Mahony for their insightful comments and feedback.

Cover image by: Alexis Verback