Automation and the Future of Jobs

By Daron Acemoglu

Posted on June 28, 2017


We are in the midst of huge, transformative changes in the labor market in many developed economies. At the center of this transformation is a wave of technologies based on the computer chip that aim at automating a range of tasks previously performed by labor. Advances in artificial intelligence and robotics are the next potentially powerful phase of this wave. Despite much discussion of automation and what it spells for the future of labor markets, we are far from both a comprehensive framework for studying how automation impacts the functioning of modern labor markets, and from a body of empirical work providing reliable estimates on its impact on employment, wages and productivity.


Image: Automated factoryThis essay provides an overview of a conceptual framework for understanding the implications of automation, and a brief discussion of some recent work on the implications of robots on the US labor market. I start with a brief recap of the canonical way in which labor economists and macroeconomists think about the effects of technologies, including computer-based ones, on inequality. I then explain why this framework is not just restrictive, but it flies in the face of several key facts of US labor markets, and even more so after the advent of automation technologies. After outlining an alternative framework and its implications for wages and employment, I move on to a brief discussion of recent evidence on the effects of one salient type of automation technology, robotics, on wages and employment.




The canonical framework used by labor and macro-economists for thinking about the effects of technology on wages and employment can be summarized as the enabling technology view. Here, new technologies are conceptualized as augmenting the capabilities of some workers and enabling them to perform new functions, increasing their productivity. Arguably the first computer ever invented, the Antikythera mechanism is an example of an enabling technology from ancient Greece around 200 BC.


This mechanism enabled skilled early astronomers to calculate the positions of stars and planets, an amazing achievement that would not have been feasible without this technology. Modern examples include computer-assisted design (CAD) machines, which increase the productivity of skilled workers and design tasks and PCs, which have become an indispensable aid to all sorts of managerial and clerical workers.


As these examples illustrate, even in the enabling technology view, new technologies will help certain types of workers more than others, and thus could impact inequality. This is in fact the key to the canonical framework for the analysis of labor market and equality first introduced by the Dutch economist Jan Tinbergen, which is then developed and fruitfully applied to data in many settings. An important implication of this framework is obtained by positing that the productivity of and the demand for high-skill workers rather than low-skill workers increases more rapidly over time, increasing the wage premium of high-skill workers. However, this tendency can be counterbalanced by an increase in the supply of high-skill workers, which is the basis of Tinbergen’s famous race between technology and supply of education. According to this perspective, skill premia and wage inequality increase when technology changes faster than the supply of skills, and contracts when supply outpaces technology.


Though this framework has been extremely useful in interpreting the broad trends in the labor market of the United States and other advanced economies, it faces at least three fundamental challenges. The first one is that, despite its early success in accounting for the changes in the college premium (average earnings of college-graduate workers relative to high school graduates), this framework has done much less well in recent times.


Second, even more critically, the enabling technology view implies that any improvement in technology should lead to higher wages for all types of workers. But wage declines for low-education workers have been the norm not the exception over the past 30 years in the US labor market. In particular, the real wages of workers with less high school, high school or some college have all fallen sharply since the early 1970s. The inability of this conical framework to account for the pervasive phenomenon of declining real wages of certain groups of workers is one of its most jarring shortcomings.


Third, a more detailed look at distributional wages shows that there are richer dynamics than those that can be explained by a framework where inequality is created by the changing rewards to a single, well-defined type of skill. In particular, wages at the bottom, median and the top move very differently over different time periods. Most notably, in contrast with simple skill-biased technological change view, we do not see an opening of the gap between median and bottom wages. Rather, following a period of sharp falls at the bottom of the wage distribution, there is an extended period from the mid-1980s to the mid-1990s where wages at the bottom are increasing more rapidly than wages in the middle of the distribution.


In contrast to a view based on enabling technologies helping the most highly skilled workers, we see rapid employment growth at the bottom of the wage distribution both in the 1990s and 2000s. The picture that emerges is thus one in which the economy is generating considerably more employment in lower-paid occupations than in occupations in the middle of the wage distribution.


Finally, we can also verify that this is not just a US phenomenon. The middle-paying occupations have contracted in every European country between 1993 and 2006, strongly suggesting that the employment patterns we are witnessing in the United States are due to common technological trends rather than idiosyncratic US factors.




The alternative to the enabling technologies view is to conceptualize new technologies as explicitly replacing labor in some tasks. Of course, in practice some technologies will be enabling, like the Antikythera mechanism or computer-assisted design technologies, while others will be replacing. The perspective in this essay is that many of the new technologies transforming the labor market are not of the enabling type but clearly replacing and displacing labor, and this has far-reaching consequences.


The classic historical example of replacing technology is the Jacquard Loom, a power loom invented in 1801, which significantly simplified intricate weaving steps in textile manufacturing. Today, various computer-based automation technologies such as automated teller machines, computerized inventory control and mail sorting machines are examples of replacing technologies. Most major replacing technologies that have already started spreading in the economy are industrial robots, which take over various tasks previously performed by semi-skilled industrial workers, and artificial intelligence, which promises to replace workers in many skilled occupations ranging from paralegals to accountants and even some middle managers.


Conceptually, we can make sense of replacing technologies by abandoning the reduced-form formalization of the relationship between technology and factors of production used above, and think instead in terms of tasks that need to be performed for production.


In addition to descriptive richness of this task-based framework, it has the advantage of providing a conceptual framework in which the challenges facing the enabling technologies view can be readily resolved. In particular, in this framework:

  • In contrast to the standard framework based on enabling technologies, replacing technologies can reduce wages. This contrasts with the predictions of the canonical model we discussed in the previous section. The key is the difference between enabling and replacing technologies. As already noted, enabling technologies, by augmenting one type of labor or the other, always increase the demand for both factors of production. This is not the case with replacing technologies. Even with a single type of labor competing against technology or capital, a set of tasks shifting from labor to capital can reduce wages. This effect is further strengthened if there are multiple types of labor, and new technologies directly take away some of the tasks performed by a specific type of labor (for example, semi-skilled manufacturing workers or operators).
  • For the same reasons as articulated in the previous bullet point, replacing technologies displace workers, and may cause unemployment.
  • If new technologies replace tasks in the middle of the pay distribution, they will cause polarization of employment. Intuitively, these new technologies will take away the middle paying occupations, and thus the overall wage distribution will have a smaller, in some sense ‘hollowed’ middle, causing wage polarization. Interestingly, because workers dislocated by technology from the middle of the pay distribution will compete with others, changes in employment structure may be divorced from wage growth patterns. As a result, we may expect to find faster growth of employment in lower-paying occupations as those dislocated by technology also seek employment in these occupations, which is confirmed by the changes in employment structure shown in the figure below, but this does not necessarily imply faster wage growth in these expanding occupations.
Image: Chart of changes in employment

It is also worth noting that the relevance of replacing technologies also stems from the fact that many of the major recent technological waves, which have included advances in automation, robotics and artificial intelligence, fit much more closely with the conceptualization of new technologies. In fact, the spread of industrial robots is a perfect case study for replacing technologies, which we turn to next.




So what do we know about the effects of automation or more specifically robots on jobs and wages? Do they tend to increase wages for all types of workers as an approach predicated on the enabling technologies view would imply? Or do they dislocate many types of workers, reducing their employment and wages as the replacing technologies view would maintain?


Despite the recent ubiquity of these types of technologies, we know surprisingly little about these questions. Most of what we know comes from studies that investigate how feasible it is to automate existing jobs given current and presumed technological advances. For instance, Frey and Osborne (2013) classify 702 occupations by how susceptible they are to automation based on the current set of tasks they perform. They conclude that over the next two decades, 47 percent of US workers are at the risk of automation. A recent report by McKinsey applies this methodology somewhat differently but arrives at similar conclusions: 45 percent of US workers are at risk of losing their jobs in the face of automation. The World Bank’s feasibility study goes even further and finds that 57 percent of jobs in OECD countries could be automated and wither away over the course of the next two decades.


But there are several reasons for not fully trusting the conclusions from these studies. First, it is notoriously difficult to estimate which jobs can be fully automated. For example, another paper utilizing the same broad methodology, Arntz, Gregory, and Zierahn (2016), reaches a very different conclusion because it maintains that within an occupation, many workers specialize in tasks that cannot be automated easily. Their conclusion is that once this type of specialization is taken into account only about 9 percent of jobs in the OECD are at risk. Second, even more fundamentally, these feasibility approaches do not take equilibrium economic responses into account. Feasibility of automating a task does not mean that firms will find it profitable to automate it. And more importantly, the full (and arguably interesting) labor market impacts of new technologies depend not only on where automation and robotics might directly impact, but also on how the rest of the economy will adjust. Most importantly, during several other episodes of major technological change (including rapid automation), other, sometimes new, sectors and occupations have expanded, keeping employment and wages high.


Recent work by Acemoglu and Restrepo (2017), in a working paper titled ‘Robots and Jobs: Evidence from US Labor Markets’, goes beyond these feasibility studies to estimate the equilibrium impact of industrial robots on jobs and wages. Industrial robots are defined by the International Federation of Robotics (IFR) as “an automatically controlled, reprogrammable, and multipurpose [machine]”. That is, industrial robots are machines that do not need a human operator and that can be programmed to perform several manual tasks such as welding, painting, assembling, handling materials, or packaging. Since 1993, industrial robots have been spreading in workplaces, with a global stock reaching more than 1.5 million today. Most experts estimate that robots will become much more ubiquitous in the next decade or so.


Acemoglu and Restrepo (2017) focus on the local labor market effects of robots. Their empirical strategy relies on a measure of change in exposure to robots, constructed using data from the IFR on the increase in robot usage among 19 industries (roughly at the level of two-digit industries) and their employment shares from the Census before the onset of recent robotic advances (in practice 1990). This measure of the change in exposure to robots captures the variation in the distribution of industrial employment across areas around 1990. The reasoning of this measure stems from a simple model of automation and effects of industrial robots, which intuitively relies on the fact that the industry-level adoption of robots in the United States will be related to other industry trends or economic conditions in commuting zones specializing in an industry, the relationship between exposure to robots and labor market outcomes could be confounded. To address this concern, Acemoglu and Restrepo (2017) use the industry-level spread of robots between 1990 and 2007 in other advanced economies, meant to proxy improvements in the world technology frontier of robots, as an instrument for industry trends in the United States. Though not a panacea for all sources of omitted variable biases, this strategy has the advantage of focusing on the variation that results solely from industries in which the use of robots has been concurrent in all or most advanced economies. Moreover, because IFR industry-level data starts only 2004 in the United States, but in 1993 in several European countries, this strategy also enables us to study the impact of industrial robots from 1990 to 2007.


Using this strategy leads to fairly precise, large and negative estimates of robots on employment and wages, very much in line with the replacing technologies view of the world. In particular, in commuting zones that have experienced the largest increase in exposure to robots, there are precisely estimated declines in employment and wages between 1990 and 2007. There are many concerns about the interpretation of these results, especially since other changes affecting local labor markets in the United States might confound the effects of robots. However, these estimates appear to be very robust to controlling for broad industry composition, for detailed demographics, and for competing factors impacting workers in commuting zones - in particular, exposure to imports from China and the decline in routine jobs following the use of software to perform information processing tasks. Perhaps as importantly, most affected commuting zones do not appear to be on a differential trend before the onset of the rising robot usage circa 1990.


Quantitatively, these estimates imply that one new robot per thousand workers reduces the US employment to population ratio by 0.18-0.34 percentage points and average wages by 0.25-0.5 percent. The employment effects are equivalent to one more robot reducing aggregate employment by about three workers, which is not implausible.


Reassuringly, it appears that the effects of robots are concentrated on the most heavily automated industries; on routine manual, non-routine manual and blue-collar occupations; and on workers with less than college education. The effects on men and women are similar, though somewhat larger on men.




Image: Robots on production lineTo understand the transformative changes our economy and labor market are undergoing, we need to depart from the canonical way in which economists think about technology – as a tide that lifts all boats. Many technologies, which this essay has called ‘replacing technologies’, displace workers by substituting machines and capital for tasks previously performed by labor. They can reduce, in the short run and the medium run, wages and employment. This makes it more critical to develop a broader approach to the adjustment of the economy in the face of new technologies, because economic adjustment left to its own devices will create considerable hardship for many workers.


After providing a brief overview of this conceptual structure and how it differs from the canonical approach in economics, this essay summarized recent work on the effect of an exemplar of this type of replacing technology, robots, on employment and wages. The evidence indicates large wage and employment losses resulting from the introduction of industrial robots into manufacturing.



The preceding is republished on TAP with permission by its author, Professor Daron Acemoglu and the Toulouse Network for Information Technology (TNIT). “Automation and the Future of Jobs” was originally published in TNIT’s May 2017 newsletter.


Read more from Professor Daron Acemoglu on Artificial Intelligence and the Labor Market:



About the Author

  • Daron Acemoglu
  • Massachusetts Institute of Technology
  • 50 Memorial Dr., E52-380b
    Cambridge, MA 02142-1347

Recent TAP Bloggers