Technology-Skill Complementarity and Labor Displacement: Evidence from Linking Two Centuries of Patents with Occupations

We construct new technology indicators using textual analysis of patent documents and occupation task descriptions that span of two centuries (1850–2010). At the industry level, improvements in technology are associated with higher labor productivity but a decline in the labor share. Exploiting variation in the extent certain technologies are related to specific occupations, we show that technological innovation has been largely associated with worse labor market outcomes— wages and employment—for incumbent workers in related occupations using a combination of public-use and confidential administrative data. Panel data on individual worker earnings reveal that less educated, older, and more highly-paid workers experience significantly greater declines in average earnings and earnings risk following related technological advances. We reconcile these facts with the standard view of technology-skill complementarity using a model that allows for skill displacement. ∗We are grateful to Daron Acemoglu, David Autor, Martin Beraja and seminar participants in the meetings of the Econometric Society and the Society of Economic Dynamics for valuable discussions and feedback, and to Will Cong for generously sharing his replication code. Brice Green and Jinpu Yang provided excellent research support. The paper has been previously circulated as “Technological Change and Occupations over the Long Run” †MIT Sloan School of Management and NBER ‡Kellogg School of Management and NBER §MIT Sloan School of Management ¶MIT Sloan School of Management Economists and workers alike have long worried about the employment prospects of workers whose key tasks can be easily performed by a machine, robot, software, or some other form of capital that substitutes for labor.1 These concerns have been exacerbated by recent breakthroughs in automation technologies (e.g., software, artificial intelligence, robotics) which have expanded the set of manual and cognitive tasks which can performed by machines and have occurred contemporaneously with an increase in income inequality and a fall in the labor share of aggregate output.2 Yet, despite the importance of these issues, systematic evidence for technological displacement remains elusive.3 Our goal is to fill this gap: we leverage over a century and a half of data to propose and validate new metrics of workers’ exposure to technological innovation and relate them to workers’ labor market outcomes, both at the aggregate as well as the individual level. To measure workers’ exposures to technical change we measure the similarity between the textual description of the tasks performed by an occupation and that of major technological breakthroughs. We identify the later through the textual analysis of patent networks using the methodology of Kelly, Papanikolaou, Seru, and Taddy (2020). To estimate the distance between a breakthrough innovation and workers’ task descriptions, we leverage recent advances in natural language processing that allow us to compute a measure of the similarity between documents that accounts for synonyms. By exploiting the timing of patent grants we can identify the extent to which certain worker groups (occupations) are exposed to major technological breakthroughs at a given point in time. In sum, our indices capture the extent to which specific occupations are exposed to breakthrough innovations in a given year. We emphasize that, a priori, we are agnostic on whether innovations that are similar to tasks certain occupations perform are likely to be substitutes or complements. For that, we need to examine how our indicators correlate with labor market outcomes. A key advantage of our methodology is that it relies only on document text; as such, we are able to construct time-series indices of occupation exposures that span the last two centuries. For example, our technology exposure for “molders, shapers, and casters, except metal and plastic”—an occupation category which includes glass blowers as a sub-occupation—takes a relatively high value in the 1Fear of technological unemployment is not new. In 350 BCE, Aristotle wrote: “[If] the shuttle would weave and the plectrum touch the lyre without a hand to guide them, chief workmen would not want servants, nor masters slaves.” In 1811, skilled weavers and textile workers (known as Luddites) worried that mechanizing manufacturing (and the unskilled laborers operating the new looms) would rob them of their means of income. In 1930, Keynes described this type of potential labor market risk when he said, “We are being afflicted with a new disease of technological unemployment...due to our discovery of means of economising the use of labor outrunning the pace at which we can find new uses for labor." More recently, a McKinsey report estimated that between 400 million and 800 million jobs could be lost worldwide due to robotic automation by the year 2030. 2For instance, one of the leading explanations for the increase in the skill premium is skill-biased technical change, whereas the decline in the labor share has been attributed to capital-embodied technical change.. See Goldin and Katz (2008); Krusell, Ohanian, Ríos-Rull, and Violante (2000); Karabarbounis and Neiman (2013); Acemoglu and Restrepo (2020, 2018, 2021) 3Due to the difficulty of constructing broad measures of labor-displacive innovations, existing work has focused on analyzing specific instances in which the impact of a specific technology on workers can be identified (Atack, Margo, and Rhode, 2019; Feigenbaum and Gross, 2020; Akerman, Gaarder, and Mogstad, 2015; Humlum, 2019).