Automation and Polarization

he research article by Daron Acemoglu and Jonas Loebbing has been published in the Journal of Political Economy!

We asked Jonas Loebbing to share their key insights.

Discover why firms may choose to automate middle-complexity tasks first, how this can produce employment and wage polarization, and why future waves of AI could change which workers are most exposed.

Automation and artificial intelligence are increasingly transforming workplaces and reshaping job tasks, while also raising questions about their effects on workers and labor markets. In your research article, you tackle exactly this topic. Could you explain what the central question of your paper is?

In the paper, we ask why automation has so often been linked to polarization of the labor market. In many countries, workers in the middle of the wage distribution have seen weaker employment and wage growth than workers at the bottom or the top. A common explanation is that middle-skill jobs contain routine tasks that are technologically easy to automate. Our article offers a complementary explanation: these tasks may also be the most profitable for firms to automate.

What do you mean by polarization of the labor market?

Polarization means that the labor market becomes more divided. Occupations in the middle of the wage distribution are squeezed, while occupations at the lower and upper ends are growing. In employment terms, workers move away from medium-pay jobs towards lower or better paid jobs. In wage terms, workers with salaries in the middle of the wage distribution lose ground relative to those below and above them.

So your argument is not only that machines can do middle-skill tasks, but that firms have a stronger incentive to automate them?

Indeed. Firms do not automate tasks simply because technology makes it possible. Instead, they automate when doing so is economically worthwhile. This creates different patterns across the labor market. At the lower end, wages are so low that replacing workers with machines often brings little cost advantage. At the higher end, tasks are so complex that human expertise still outperforms current technology, despite higher wages. The most exposed group is the middle: wages are high enough to make automation attractive, while the tasks are sufficiently feasible to automate.

How does the model capture this idea?

The model imagines production as a set of many tasks, ranging from simple to complex. Workers also differ by skill. Without automation, lower-skill workers do simpler tasks and higher-skill workers do more complex tasks. The new element is that capital - machines, software, robots, or algorithms - can also perform some tasks. The key question is which tasks firms choose to assign to capital rather than labor.

And what do you find?

The main result is that automation typically starts in the middle of the task distribution. We call this interior automation. When that happens, a further decline in the cost of capital pushes workers away from the middle and toward the two ends of the task distribution. This creates employment polarization. It also creates wage polarization: workers near the newly automated tasks see their wages decline relative to those at the bottom and the top of the distribution.

That sounds counterintuitive. Why would low wages at the bottom make middle-skill workers more vulnerable?

Low wages at the bottom make it cheap to keep using human labor for low-complexity tasks. If a firm can hire workers cheaply for those tasks, the machine does not save much money. Middle-skill workers are paid more, so replacing them can generate larger savings. In this sense, low wages at the bottom can shelter some low-skill tasks from automation, while making the middle more exposed.

Does automation always reduce wages?

No, though it often reduces them for specific groups of workers. Whether a worker's wage falls depends on the balance between two competing forces: the displacement effect and the productivity effect.

Could you elaborate more on these two forces?

The displacement effect means that when machines or algorithms take over tasks previously performed by workers, demand for those specific tasks falls, which can put downward pressure on wages. At the same time, there is a productivity effect. Automation reduces production costs and thereby raises the demand for the goods produced. This effect increases the demand for all tasks that are not automated, supporting higher wages, especially for workers whose tasks are not directly affected by automation.

Where does the quantitative part of your paper come in, and what does it focus on?

Our quantitative analysis focuses on the United States, where detailed data allow us to examine how automation has affected the labor market over recent decades.

We calibrate the model to reproduce some salient changes in the US wage distribution from 1980 to 2016–17.

We then use the model to examine three counterfactual scenarios: another wave of automation, an AI-like increase in capital productivity, and the introduction of a $16 minimum wage.

And what do you find in case when automation becomes much more powerful, for example through AI?

As AI becomes more powerful, it changes who is most exposed to automation. Automation can become cheap and capable enough to replace even low-paid work, expanding its reach down the skill ladder. Rather than only hollowing out the middle, the impact can become more linear: the lower the wage, the higher the risk of displacement. High-skill jobs remain the least affected and may even benefit from increased productivity through AI. Yet, we acknowledge that the future development of the capabilities of AI is hard to predict and we consider this part of our paper more speculative than our other simulations.

And what do you find in the context of minimum wage?

Minimum wages raise the cost of employing low-skill workers. That can make low-complexity tasks more attractive to automate. In the model, a sufficiently high wage floor makes interior automation less likely and can push the economy toward low-skill automation. This does not mean minimum wages are bad policy. It means they can change firms' automation incentives, so policy design needs to consider both wages and technology.

What is the broader policy message?

The effects of automation on workers are not determined by technology alone, but also by institutional choices and by the current situation on the labor market. A central implication is that policymakers should be aware of a potential shift in automation from middle-skill jobs toward low-skill work. In that context, well-intended policies such as higher minimum wages can have unintended consequences. By raising the cost of low-wage labor, they may accelerate incentives for firms to automate exactly those jobs.

Yet, in our view, the main message is not to refrain from such policies but to accompany them with policies that encourage investments into more labor-complementary technologies instead of further automation. One way to achieve this would be to strengthen the role of labor unions in firms. Empirical evidence shows that firms with strong labor unions are more likely to invest into technologies that support, rather than replace, their workers.

Read the article:

Daron Acemoglu and Jonas Loebbing: Automation and Polarization in Journal of Political Economy.