Why AI Advancements May Push Some Workers Out of the Labor Force

Recent discussion of AI technology’s effects on labor markets has focused on unemployment. The rising unemployment rates in occupations with higher AI prevalence speak to the short-run effects of uncertainty in the face of new technology.

But how might one envision the possible effects of improvements in AI technology in the long run? What does Econ 101 tell us? The aim of this blog post is to offer simple tools of economic analysis to help think through the likely consequences of AI adoption in the long run. The goal is not to predict the future, as this is a complicated question, but rather to clarify first-order economic forces at play.

I use a simple model of comparative advantage to show the likely consequences of increased productivity from AI technology. In my analysis, AI is viewed as complementing skilled workers and making them incredibly productive across a broad range of economic activity. Conventional jobs, however, require skilled workers for supervision and other professional support. As the value of skilled time rises with advances in AI technology, it becomes optimal in this model economy to eliminate conventional jobs so as to allow the skilled labor previously engaged in supervision to engage in more productive activity.

The key findings that emerge from this simple setup suggest that large improvements in AI technology:

• Will increase total output and average labor productivity, and therefore can raise consumption for everyone, depending on how goods and services are distributed across workers
• Will expand employment of AI-using workers into more sectors
• May push some workers—those relying on conventional technology—entirely out of employment

The findings suggest that workers who lose out from AI advancement could benefit from having a safety net that helps them share consumption gains with AI-using workers. In addition, retraining programs could ensure more workers have the skills to harness productivity gains from effective use of AI.

A Simple Comparative Advantage Economy

Sectors and Workers

Consider a model economy with two sectors of production: apples and computers. There are two types of workers:

• 100 skilled workers, who effectively use AI at work
• 40 unskilled workers, who rely on conventional methods of production

The “skilled” label refers to worker effectiveness at harnessing productivity gains associated with AI. In practice, even highly educated workers in professional jobs (e.g., a data analyst) may fall into the “unskilled” category if they do not effectively engage with AI.

Technology

Workers can work in either sector but differ in how productive they are in each. Specifically, I assume one skilled worker can produce either 32 computers or 12 apples, while one unskilled worker can produce either 0 computers or 9 apples.

I also assume unskilled workers require support from skilled workers. This support requirement is meant to represent supervision, IT services and other professional inputs. Specifically, each unskilled worker requires one-fourth of a skilled worker’s time. This assumption is critical for our main result.

Demand Factors

Suppose workers like to consume a balanced basket of goods, so that the demand for apples and computers remains proportional. For the sake of getting round numbers, let us assume consumers demand eight computers for every three apples they demand.I omit price determination from this discussion, but prices would stay fixed across the two cases that were studied.

Baseline Economy: How to Best Allocate AI-Using Workers and Others across Sectors

In this baseline economy, how should workers be allocated across sectors to ensure the computers-to-apples ratio is 8-to-3 while maximizing production?

The answer is given in the first row of the table below (the “baseline” case). Because skilled workers are relatively more productive in making computers, it is optimal to use them in computer production, while the opposite is true for the unskilled workers. If one were to allocate the 40 unskilled workers to making apples, 10 of the skilled workers to supervising the unskilled workers and the remaining 90 skilled workers to making computers—the case of complete specialization—the economy would produce 360 apples (40 x 9) and 2,880 computers (90 x 32). This perfectly specialized labor allocation results in too many computers per apple based on consumers’ demand for a more balanced basket of goods. Therefore, it is optimal to allocate some of the 90 skilled workers to producing apples. We do so until we reach the desired balance of goods. Allocating 30 skilled workers to apple production yields the desired balance: 720 apples and 1,920 computers.

Advanced AI: What Happens to Labor Reallocation When AI Gets Better?

We think of improvements in AI technology as making the AI-using workers more productive in both sectors. What happens in this model when we quadruple AI productivity so that one skilled worker can produce either 128 computers or 48 apples?

As before, the skilled workers will engage in production of both goods. But does it make sense for the unskilled workers to still engage in apple production? One unskilled worker produces nine apples but also requires one-fourth of a skilled worker’s time in professional support and supervision. However, during that time, a skilled worker could produce 12 apples (0.25 x 48). Thus, the economy can increase the production of apples by eliminating the unskilled jobs and reallocating skilled workers previously used in professional support and supervision to apple production.

The second row of data in the table above outlines the optimal labor allocation in this economy. Skilled workers are equally distributed between the two sectors of production, which yields the desired 8-to-3 ratio of computers to apples. The economy produces more than ever, but unskilled employment falls to zero.

Key Takeaways

This simple model delivers two important lessons. First, AI raises total output—and it has the potential to raise consumption for everyone, depending on whether redistribution policies are in place that allow the gains from technological progress to be broadly shared. Second, a sufficiently large improvement in AI technology can make some forms of labor obsolete, which highlights the importance of investing in skills that help workers use AI effectively.

Concluding Remarks

The impact of advanced AI will likely differ from the effect of “machine” automation of the past. When complemented with skilled human input, AI has the potential to contribute to a broader range of economic activity, which is what our simple model aimed to capture. This view of AI also implies that the financial returns to AI advances will tend to accrue to the highly skilled human input rather than to capital as in the case of automation, and that these returns will therefore be shared more unevenly across the economy.

Nonetheless, there are goods and services that will be valued for the human input per se (e.g., art, music, social services, competitive sports). As the economy grows, the demand for these goods and services will also rise and new jobs will be created. In this Massachusetts Institute of Technology report on the work of the future (PDF), David Autor, David Mindell and Elisabeth Reynolds gave an overview of job evolution through the era of automation. They concluded: “No economic law dictates that the creation of new work must equal or exceed the elimination of old work. Still, history shows that they tend to evolve together.”

Note

1. I omit price determination from this discussion, but prices would stay fixed across the two cases that were studied.