Is AI Creating or Destroying Jobs?

By Daron Acemoglu

Posted on October 21, 2021

Graphic image of human shaking hands with a robot who is standing in front of a smartphone.

Artificial intelligence (AI) is one of the most promising technologies currently being developed and deployed. There is a lot of excitement, some hype, and a fair bit of apprehension about what AI will mean for our security, society, and economy.


Central to many of the debates is whether AI is creating or destroying jobs. Despite the huge and growing interest in this question, we know relatively little about the answers.


Some commentators are convinced that AI is the harbinger of a jobless future (e.g., Ford, 2015; West, 2018; Susskind, 2020). Yet others are equally adamant that AI will enrich work experiences and increase human productivity, contributing rather than detracting from job growth (e.g., McKinsey Global Institute, 2017). These contrasting visions persist in part because there is very little evidence on what AI is doing to work and workers. There are currently no representative data sets of AI, so we lack representative evidence on whether there has been a major increase in AI adoption (as opposed to just talk of AI). It is possible to find examples of AI technologies either replacing work or complementing workers, precisely because AI, as a broad technological platform, is capable of doing both. The level of job displacement that AI will create is thus partly a matter of societal and business choice (Acemoglu and Restrepo, 2019).


Tracking the Rise of AI Activity


In recent work, “AI and Jobs: Evidence from Online Vacancies”, David Autor, Jonathon Hazel, Pascual Restrepo and I have studied AI adoption in the US and its labor market implications. AI adoption can be partially identified from the footprint it leaves at adopting establishments as they hire workers specializing in AI-related activities, such as supervised and unsupervised learning, natural language processing, machine translation, or image recognition. To put this idea into practice, we built an establishment-level data set of AI activity based on the near-universe of US online job vacancy postings from Burning Glass Technologies for the years 2007 and 2010 through 2018. This data set, which has been used in several recent papers, contains detailed information on occupations and the skills required for each posted vacancy.


“There are still relatively few vacancies in core AI areas, such as machine learning and natural language processing, though the rate of growth since 2015-2016 has been staggering.”


We then linked the adoption of AI and its possible implications to the task structure of an establishment. Put simply, the idea is to look for a set of “AI-suitable” tasks, which may be targeted by AI applications, then investigate whether establishments with a high fraction of such tasks are more likely to show rapid AI adoption, as measured by the hiring of AI workers. There is no consensus on which tasks are AI-suitable. Nevertheless, a number of recent studies have started developing systematic ways of measuring which tasks can be performed or augmented by current AI technologies. For example, Felten, Raj and Seamans (2018) construct an index of the effect of AI on various occupations, meant to capture both the ability of AI algorithms to substitute for humans and their complementarity to humans. They build on experts’ assessments of areas in which AI has made important advances then map these areas to the set of tasks performed by different occupations. Alternatively, Brynjolfsson, Mitchell and Rock (2018) build a measure of the suitability of an occupation’s task to be performed by machine learning. Webb (2000), on the other hand, uses natural language processing on the text of patents to map them to specific tasks performed within various occupations. Each of these measures captures a different aspect of AI suitability (and we show they are quite distinct). There is information in each of them and our work uses all three of them.


The data paint a clear picture about AI activity, regardless of which specific measure one looks at. There is a notable takeoff in AI vacancy postings starting in 2010, but these postings remain very low until around 2015-2016, then undergo an inflection, trending up strongly thereafter. There are still relatively few vacancies in core AI areas (such as machine learning, natural language processing, etc.), though the rate of growth since 2015-2016 has been staggering.


“The fact that AI is a broad technological platform suggests there are important decisions for both corporations and public policymakers. What type of AI do we want? Can we create more jobs than we destroy?”


We also find that AI-adopting establishments start demanding different skills than before, and in fact there is some evidence of increased “skills churn” associated with AI. This bolsters the case that there are significant changes in the organization of production and thus the skills demanded by businesses at the forefront of AI adoption.


Where Does this Growth Come From?


We show that there is a strong association between the baseline task structure of an establishment and AI activity. This relationship is present with all of the measures mentioned above. As important, it remains even when we focus on establishments within a narrow industry or, more notably, when we compare two establishments belonging to the same multi-establishment firm that still differ in terms of their baseline task structure. This is evidence that AI adoption is being at least partially targeted to a specific set of AI-suitable tasks. This correlation, however, does not answer the key question we started with …


Is AI Creating or Destroying Jobs?


The answer seems to be: Mostly, it’s too soon to tell. Despite a remarkable takeoff, there is still very little AI activity at the moment, and AI-impacted job changes may be a small drop in a big bucket. The number of AI-suitable tasks may grow and lead to the hiring of more workers than before because of the rollout of AI technologies. Or, conversely, the workers who previously performed these tasks may be replaced by AI algorithms.


Nevertheless, we see some evidence of fewer vacancies for non-AI positions in the more heavily impacted establishments (those with a high fraction of AI-suitable tasks). For example, establishments with a high share of AI-suitable tasks in 2010 subsequently show significantly slower growth in vacancies. Yet, confirming our conclusion that AI activity is still too small, this establishment-level result does not translate into slower growth in the more AI-exposed occupations or industries.


So where do we go from here?


The evidence we gathered – coupled with advances in machine learning, big data and other areas of AI – suggests that the rapid takeoff in AI activity will continue in the years to come. This may imply more displacement (similar to the negative hiring effects we may be seeing already at some establishments), but AI is a broad technological platform and can be used in many different ways. The fact that early AI is targeted to specific tasks does not mean that, as the technology matures, it will not have other applications. There is already evidence that AI technologies are being used for new product development and reorganization (Bresnahan, 2019), and these uses may intensify in the years to come.


The fact that AI is a broad technological platform also suggests that there are important decisions for both corporations and public policymakers. What type of AI do we want? If AI can create and destroy jobs at the same time, can we make sure that we create more jobs than we destroy?


We sometimes hear a narrative suggesting that there is a clear path of future technology. For AI, a broad technological platform with many applications, this may be particularly untrue. The disagreement about the effects of AI for workers is rooted in the fact that AI can destroy as well as create jobs. But this also implies there is a lot of room for public policy and corporate strategies in shepherding AI in a direction that is more beneficial for society.


Key Findings

Graphic image of a large brain inside a container that has computers connected to it. Three people are looking at monitors connected to the container.
  • AI vacancy postings took off in 2010, but remained very low until around 2015-2016, trending up strongly thereafter.
  • There are still relatively few vacancies in core AI areas but the rate of growth has been staggering and looks set to continue.
  • We see some evidence of fewer non-AI vacancies in establishments heavily impacted by AI. Yet this does not yet translate into slower growth in the more AI-exposed occupations or industries.
  • We also detect the beginning of an AI-driven “skills churn” in which the use of AI technologies is associated with changes in the skills demanded.

Further Reading


For more research by Daron on this subject, see our previous edition of TNIT News (Issue 17):


Daron Acemoglu is an Institute Professor at MIT and is the author of five books. His academic work covers a wide range of areas, including political economy, economic development, economic growth, technological change, inequality, labor economics, and economics of networks. He was awarded the John Bates Clark Medal in 2005, the Erwin Plein Nemmers Prize in 2012, and the 2016 BBVA Frontiers of Knowledge Award.


Daron has received the inaugural T. W. Shultz Prize from the University of Chicago in 2004, and the inaugural Sherwin Rosen Award for outstanding contribution to labor economics in 2004, Distinguished Science Award from the Turkish Sciences Association in 2006, the John von Neumann Award, Rajk College, Budapest in 2007, the Carnegie Fellowship in 2017, the Jean-Jacques Laffont Prize in 2018, the Global Economy Prize in 2019, and the CME Mathematical and Statistical Research Institute Prize in 2021.


The preceding is republished on TAP with permission by its author, Professor Daron Acemoglu, and by the Toulouse Network for Information Technology (TNIT). “Is AI Creating or Destroying Jobs?” was originally published in TNIT’s September 2021 newsletter.