Assistant Professor of Operations,
Information, and Decisions
Wharton School, University of Pennsylvania
Publications and Working Papers
What Can Machines Learn, and What Does It Mean for Occupations and the Economy? (with Erik Brynjolfsson and Tom Mitchell)
(published in AEA: Papers and Proceedings 2018)
Advances in machine learning (ML) are poised to transform numerous occupations and industries. This raises the question of which tasks will be most affected by ML. We apply the rubric evaluating task potential for ML in Brynjolfsson and Mitchell (2017) to build measures of "Suitability for Machine Learning" (SML) and apply it to 18,156 tasks in O*NET. We find that (i) ML affects different occupations than earlier automation waves; (ii) most occupations include at least some SML tasks; (iii) few occupations are fully automatable using ML; and (iv) realizing the potential of ML usually requires redesign of job task content.
We live in an age of paradox. Systems using artificial intelligence match or surpass human-level performance in more and more domains, leveraging rapid advances in other technologies and driving soaring stock prices. Yet measured productivity growth has declined by half over the past decade, and real income has stagnated since the late 1990s for a majority of Americans. We describe four potential explanations for this clash of expectations and statistics: false hopes, mismeasurement, redistribution and implementation lags. While a case can be made for each explanation, we argue that lags have likely been the biggest contributor to the paradox. The most impressive capabilities of AI, particularly those based on machine learning, have not yet diffused widely. More importantly, like other general purpose technologies, their full effects won’t be realized until waves of complementary innovations are developed and implemented. The adjustment costs, organizational changes, and new skills needed for successful AI can be modeled as a kind of intangible capital. A portion of the value of this intangible capital is already reflected in the market value of firms. However, going forward, national statistics could fail to measure the full benefits of the new technologies and some may even have the wrong sign.
GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models (With Tyna Eloundou, Sam Manning, and Pamela Mishkin)
We investigate the potential impact of large language models (LLMs) on the U.S. labor market. We develop and leverage a new rubric, assessing occupations based on their alignment with LLM capabilities, using both human and GPT-4 ratings. Our findings reveal that around 80% of the U.S. workforce have at least 10% of their work tasks exposed to the capabilities of LLMs, and approximately 19% of workers have at least 50% of their tasks exposed. We do not make predictions about the development or adoption timeline of such LLMs. LLM impact potential is pervasive, LLMs improve over time, and complementary investments will be necessary to unlock their full potential. This suggests LLMs are general-purpose technologies (Bresnahan and Trajtenberg 1995). As such, LLMs could have considerable economic, societal, and policy implications.
The Productivity J-Curve: How Intangibles Complement General Purpose Technologies (with Erik Brynjolfsson and Chad Syverson)
(published in AEJ: Macroeconomics)
General purpose technologies (GPTs) such as AI enable and require significant complementary investments,including co-invention of new processes, products, business models and human capital. These complementary investments are often intangible and poorly measured in the national accounts, even when they create valuable assets for the firm. We develop a model that shows how this leads to an underestimation of productivity growth in the early years of a new GPT, and how later, when the benefits of intangible investments are harvested, productivity growth will be overestimated. Our model generates a Productivity J-Curve that can explain the productivity slowdowns often accompanying the advent of GPTs, as well as the increase in productivity later. We use our model to analyze empirically the historical roles of intangibles tied to R&D, software, and computer hardware. We find substantial and ongoing Productivity J-Curve effects for software in particular and computer hardware to a lesser extent. Our adjusted measure TFP is 11.3% higher than official measures at the end of 2004, and 15.9% higher than official measures at the end of 2017. We then assess how AI-related intangible capital may be currently affecting measured productivity and find the effects are small but growing.
Engineering Value: The Returns to Technological Talent and Investments in Artificial Intelligence (Working Paper)
Employees with technological skills are highly complementary to the intangible knowledge assets that firms accumulate. This paper describes how technical talent is a source of rents for corporate employers, particularly for the case of Google’s surprising open source launch of TensorFlow, a deep learning software package. First, I present a simple model of how employers intangible assets expose them to the returns to their employees’ skill acquisition efforts. Then, using over 180 million position records and over 52 million skill records from LinkedIn, I build a panel of firm-level skills to measure the market value of exposure to newly available deep learning talent. AI skills are strongly correlated with market value, though variation in AI skills from 2014-2017 does not explain contemporaneous revenue productivity within firms. AI-intensive companies rapidly gained market value following the launch of TensorFlow, while companies with opportunities to automate relatively larger quantities of labor with machine learning did not. Using a difference-in-differences approach, I show that the TensorFlow launch is associated with an approximate market value increase of $11 million per 1 percent increase in AI skills for AI-using firms. AI superstar firms in the top quintile also appear to benefit more over the sample period than less intensive adopters. These results suggest that the primary mechanism responsible for the market valuation increases of AI adopters at the time of the AI skill shock from TensorFlow is a revaluation of existing firm-specific technology-exposed assets.
Digital Capital and Superstar Firms (with Prasanna Tambe, Lorin Hitt, and Erik Brynjolfsson) (Working Paper)
General purpose technologies like information technology typically require complementary firm-specific human and organizational investments to create value. These complementary investments produce a form of capital, which we call Information technology-related intangible capital (``ITIC''). An understanding of how the accumulation of ITIC contributes to economic growth and differences among firms has been hindered by the lack of measures of the stock of ITIC. We use a new, extended firm-level panel on IT labor investments along with Hall’s Quantity Revelation Theorem to construct measures of both the prices and quantities of ITIC over the last thirty years. We find that 1) prices vary significantly for ITIC, 2) significant quantities of ITIC have been accumulating since the 1990s, with ITIC accounting for at least 25% of firms’ assets by the end of our panel, 3) that it has disproportionately accumulated in small subset of high-value, superstar firms, and 4) that the accumulation of ITIC predicts future productivity.
COVID-19 and Remote Work: An Early Look at US Data (with Erik Brynjolfsson, John J. Horton, Adam Ozimek, Garima Sharma, and Hong-Yi Tu Ye) (Working Paper)
We report the results of a survey on remote work for nationally-representative sample of the US population during the COVID-19 pandemic. The survey ran in three waves in April 2020, May 2020, and July 2020 covering a total of 75,000 respondents. Of those employed pre-COVID-19, we find that about half are now working from home, including 33.0% who report they had previously been commuting and recently switched to working from home. In addition, 10.1% report being laid-off or furloughed since the start of COVID-19. We find that the share of people switching to remote work can be predicted by the incidence of COVID-19 and that younger people were more likely to switch to remote work. Furthermore, states with a higher share of employment in information work including management, professional and related occupations were more likely to shift toward working from home and had fewer people continuing to commute. We find no substantial change in results between the first two waves, suggesting that most changes to remote work manifested by early April. However, by the third wave in July, employees started to return to workplaces, with 22 percent of those who had initially switched to remote work having switched back to commuting.
Do Labor Demand Shifts Occur Within Firms or Across Them? Non-Routine-Biased Technological Change from 2000-2016 (with Seth Benzell and Guillermo Lagarda) (Working Paper)
A large literature has documented occupational shifts in the US away from routine intensive tasks. Theories of skill-biased technological change differ in whether they predict changes in occupational mix within firms, or merely across different firms or industries. Using LinkedIn resume records, BLS OES data, and Compustat employee counts, we estimate occupational employment for publicly traded US firms from 2000 through 2016. We find that faster employment growth among firms that disproportionately employ non-routine workers is the most important cause of SBTC, followed by within firm occupational mix rebalancing. The entry of new firms also plays a role, although firm exit is slightly routine-worker biased. R&D leads firms to have a larger share of routine workers. These results are most consistent with a theory of routine task demand reduction caused by the diffusion of infra-marginally implemented new technologies. We also introduce a new measure of business labor dynamism, capturing the frequency with which firms change their occupational mix. Consistent with trends in productivity and other measures of business and labor market dynamism, this measure has decreased steadily since 2000.
Identification of Peer Effects in Networked Panel Data (with Sinan Aral and Sean J. Taylor)
(published in Proceedings of the International Conference on Information Systems 2016)
After product adoption, consumers make decisions about continued use. These choices can be influenced by peer decisions in networks, but identifying causal peer influence effects is challenging. Correlations in peer behavior may be driven by correlated effects, exogenous consumer and peer characteristics, or endogenous peer effects of behavior (Manski 1993). Extending the work of Bramoullé et al. (2009), we apply proofs of peer effect identification in networks under a set of exogeneity assumptions for the panel data case. With engagement data for Yahoo Go, a mobile application, we use the network topology of application users in an instrumental variables setup to estimate usage peer effects, comparing a variety of regression models. We find this type of analysis may be useful for ruling out endogenous peer effects as a driver of behavior. Omitted variables and violation of exogeneity assumptions can bias regression coefficients toward finding statistically significant peer effects.