This conference is supported by Grant #G-2018-10104 from the Alfred P. Sloan Foundation
Hanson examines the specialization of US commuting zones in AI-related occupations over the 2000 to 2018 period. He defines AI-related jobs based on keywords in Census occupational titles. Using the approach in Lin (2011) to identify new work, the researcher measures job growth related to AI by weighting employment growth in AI-related occupations by the share of job titles in these occupations that were added after 1990. Overall, regional specialization in AI-related activities mirrors that of regional specialization in IT. However, foreign-born and native-born workers within the sector tend to cluster in different locations. Whereas specialization of the foreign-born in AI-related jobs is strongest in high-tech hubs with a preponderance of private-sector employment, native-born specialization in AI-related jobs is strongest in centers for military and space-related research. Nationally, foreign-born workers account for 55% of job growth in AI-related occupations since 2000. In regression analysis, Hanson finds that US commuting zones exposed to a larger increases in the supply of college-educated immigrants became more specialized in AI-related occupations and that this increased specialization was due entirely to the employment of the foreign born. His results suggest that access to highly skilled workers may constrain AI-related job growth and that immigration of the college-educated may help relax this constraint.
This paper was distributed as Working Paper 28671, where an updated version may be available.
Bessen, Cockburn, and Hunt investigate whether posted vacancies for jobs requiring Artificial Intelligence (AI) skills grow more slowly in U.S. locations farther from AI innovation hotspots. To define hotspots, the researchers create a geocoded dataset of all AI publications (journal articles, conference proceedings and patents) through 2020, while the researchers obtain the job vacancy information from online job advertisements scraped by Burning Glass Technologies from 2007-2019. Bessen, Cockburn, and Hunt define hotspots based on the cumulative number of AI publications by 2006. They find that a hotspot's AI publications increasingly affect other commuting zones' AI vacancies as the hotspot threshold grows to 300, a threshold met by 11% of commuting zones. A 10% greater distance from such a hotspot (about a standard deviation) reduces a commuting zone's growth in AI jobs' share of job advertisements by 2-3% of median growth. The effect is almost entirely due to the one third of job advertisements posted by employment agencies, and for which the industry is not known. A small fraction of the effect is due to job advertisements in the finance industry. The results suggests that for a minority of firms, distance from innovation is a moderate barrier to the adoption of technology.
This paper documents growing demand for worker decision-making and explores the consequences for life-cycle earnings. Career earnings growth in the U.S. has doubled since 1960, with the age of peak earnings increasing from late 30s to mid-50s. Much of this shift is explained by increased employment in decision-intensive occupations, which have longer periods of earnings growth. To understand these patterns, Deming develops a model that nests decision-making in a standard human capital framework. Work experience improves decision-making but accumulates more slowly in non-routine jobs. Life-cycle earnings in decision-intensive occupations have increased over time, with greater increases for highly-skilled workers.
This paper was distributed as Working Paper 28733, where an updated version may be available.
Copestake, Pople, and Stapleton examine the impact of artificial intelligence (AI) on hiring and wages in service sector firms, using a novel dataset of vacancy posts from India’s largest jobs website. They first document a rapid rise in demand for machine learning (ML) skills since 2016, particularly in the IT, finance and professional services industries. Vacancies requiring ML skills list substantially higher wages, but require more education and are highly concentrated both geographically and in the largest firms. Exploiting plausibly exogenous variation in exposure to advances in AI capabilities, Copestake, Pople, and Stapleton then examine the impacts of establishment demand for ML skills, as a proxy for AI adoption. They find that growth in the demand for ML skills has a direct negative impact on the total number of vacancies posted by incumbent firms. Drawing on rich data on wage offers, the researchers further find that growth in ML demand reduces wage offers for all but the lowest-paid roles.
An AI analyst Cao, Jiang, Wang, and Yang build to digest corporate financial information, qualitative disclosure, and macroeconomic indicators is able to beat the majority of human analysts in stock price forecasts and generate excess returns compared to following human analysts. In the contest of "man vs machine," the relative advantage of the AI Analyst is stronger when the firm is complex, and when information is high-dimensional, transparent and voluminous. Human analysts remain competitive when critical information requires institutional knowledge (such as the nature of intangible assets). The edge of the AI over human analysts declines over time when analysts gain access to alternative data and to in-house AI resources. Combining AI's computational power and the human art of understanding soft information produces the highest potential in generating accurate forecasts. The researchers' paper portraits a future of "machine plus human" (instead of human displacement) in high-skill professions.
This paper was distributed as Working Paper 28800, where an updated version may be available.
The following paper contains one of the first empirical studies on the firm determinants of AI adoption. The analysis relies on novel firm-level data on AI use by application and source for businesses in South Korea from 2017 and 2018. The econometric assessment identifies several firm characteristics important for AI use, specifically firm size and use of intangible assets. These characteristics are significant for AI adoption regardless of the technology’s source (i.e., produced in-house or obtained from a vendor) with some heterogeneity across different operational applications (i.e., product/service development, sales, organizational management, and so on). External partnerships are important facilitators of firm AI adoption, with some evidence for joint ventures with foreign firms in different sectors. Furthermore, AI is adopted in tandem with bundles of other digital technologies including big data, cloud computing, and the Internet of Things that facilitate data collection, usage, and processing. Finally, AI adoption corresponds with contemporaneous firm reorganization, through geographic relocation of the firm.
Calvano, Calzolari, Denicolò, and Pastorello explore the impact of algorithms that use personal data to predict tastes and make recommendations, on consumption choices and in particular on the level of market concentration. They model consumer preferences in a flexible way that allows for varying degrees of vertical and horizontal product differentiation. The analysis confirms that these algorithms produce a significant increase in market concentration even with well-behaved models of demand and against well defined economic benchmarks. The paper goes on tackling the causes and consequences of such increase. It shows that the increase is not due to the "feedback loop" created by the endogeneity of the data. Moreover, Calvano, Calzolari, Denicolò, and Pastorello show that adopting these algorithms leads to lower equilibrium prices and higher consumer surplus despite the increased level of concentration. The researchers discuss the implications of these findings for competition policy.
How does AI improve human decision-making? Answering this question is challenging because it is difficult to assess the quality of each decision and to disentangle AI's influence on decisions. Choi, Kim, Kim, and Kang study professional Go games, which provide a unique opportunity to overcome such challenges. In 2016 an AI-powered Go program (APG) unexpectedly beat the best human player, surpassing the best human knowledge and skills accumulated over thousands of years. To investigate the impact of APGs, the researchers compare human moves to AI's superior solutions, before and after the initial public release of an APG. Their analysis of 750,990 moves in 25,033 games by 1,242 professional players reveals that APGs significantly improved the quality of the players' moves as measured by the changes in winning probability with each move. Choi, Kim, Kim, and Kang also show that the key mechanisms are reductions in the number of human errors and in the magnitude of the most critical mistake during the game. Interestingly, the improvement is most prominent in the early stage of a game when uncertainty is higher. Further, young players--who are more open to and better able to utilize APG--benefit more than senior players, suggesting generational inequality in AI adoption and utilization.
of Algorithm Design
The behavior of artificial intelligence algorithms (AIAs) is shaped by how they learn about their environment. Asker, Pakes, and Fershtman compare the prices generated by AIAs that use different learning protocols when there is market interaction. Asynchronous learning occurs when the AIA only learns about the return from the action it took. Synchronous learning occurs when the AIA conducts counterfactuals to learn about the returns it would have earned had it taken an alternative action. The two lead to markedly different market prices. When future profits are not given positive weight by the AIA, synchronous updating leads to competitive pricing, while asynchronous can lead to pricing close to monopoly levels. Asker, Pakes, and Fershtman investigate how this result varies when either counterfactuals can only be calculated imperfectly and/or when the AIA places a weight on future profits.
Chen, Balasubramanian, and Forman investigate how worker mobility influences the adoption of a new general-purpose technology (GPT). Using data from over 153,000 establishments between 2010 and 2018, the researchers observe establishment decisions to adopt machine learning. Taking advantage of state-level changes to the enforceability of noncompete agreements as an exogenous shock to worker mobility, Chen, Balasubramanian, and Forman find that changes that facilitate worker movements are associated with a significant decline in the likelihood of adoption. Moreover, the magnitude of establishment response depends upon characteristics of the establishment, the location in which it resides and its industry--in particular, establishment size and number of large establishments in the same industry-location and the level of experimentation with analytics technology. These results are consistent with the view that increases in worker mobility lead to greater risks for establishments that are contemplating adoption of a new GPT that involves significant downstream innovation.
Firms increasingly rely on predictive analytics via machine learning algorithms to drive a wide array of managerial decisions. In this paper, Feng, Gradwohl, Hartline, Johnsen, and Nekipelov study the effect of competition on the choice of such algorithms, focusing on the tradeoffs between bias and variance in the algorithms' predictions. Absent competition, firms care only about the magnitude of predictive error and not its source. With competition, however, firms prefer to incur error caused by variance over error caused by bias, even at the cost of higher total error.
Aghion, Antonin, Bunel, and Jaravel use comprehensive micro data in the French manufacturing sector between 1994 and 2015 to document the effects of automation technologies on employment, sales, prices, and the labor share. Causal effects are estimated with event studies and a shift-share IV design leveraging pre-determined supply linkages and productivity shocks across foreign suppliers of industrial equipment. At all levels of analysis -- plant, firm, and industry -- the estimated impact of automation on employment is positive, even for unskilled industrial workers. They also find that automation leads to higher sales and lower consumer prices. The estimated elasticity of employment to automation is +0.31, compared with elasticities of +0.40 for sales, and -0.28 for prices. Aghion, Antonin, Bunel, and Jaravel cannot reject that the share of labor in total value added remains unchanged. Consistent with the importance of business-stealing across countries, the industry-level employment response to automation appears to be stronger in industries that face international competition. These estimates can be accounted for in a simple monopolistic competition model: firms that automate more increase their profits but pass through some of the productivity gains to consumers, inducing higher scale and higher employment. The results indicate that automation can increase labor demand and can generate productivity gains that are broadly shared across workers, consumers and firm owners. In a globalized world, attempts to curb domestic automation in order to protect domestic employment may be self-defeating due to foreign competition.
Jin and Sun collaborate with a large e-commerce platform, and conduct a year-long experiment among new sellers on the platform. The treatment group receives access to a free online entrepreneur training program customized by an AI algorithm. The researchers show that new sellers that are eligible for training see 1.7% higher revenue (intent-to-treat effect), largely driven by enhanced marketing and more efficient customer service. To investigate the mechanisms and consumer-side benefit, Jin and Sun construct a panel dataset of consumer-seller pairs across consideration sets. Using exhaustive controls and fixed effects, they first show that treated sellers have higher unobserved quality, and training expands new sellers' market share. Moreover, this effect is not driven by intensified selection through heightened entry barrier but due to a direct effect on seller quality. Jin and Sun then estimate an empirical model to capture unobserved consumer preference heterogeneity and compute welfare. Although only 0.25% of all sellers on the platform are trained, removing the program would have reduced consumer surplus by 0.07%.
The COVID-19 pandemic has devastated many low- and middle-income countries (LMICs), causing widespread food insecurity and a sharp decline in living standards1. In response to this crisis, governments and humanitarian organizations around the world have mobilized targeted social assistance programs2. A central challenge in the administration of these programs is targeting: how to identify the individuals and families with the greatest need3? This challenge is particularly acute in the large number LMICs that lack recent and comprehensive data on household income and wealth4-6. Here we show that non-traditional "big" data from satellites and mobile phone networks can improve the targeting of anti-poverty programs. Our approach uses traditional survey-based measures of consumption and wealth to train machine learning algorithms to recognize patterns of poverty in non-traditional data; the trained algorithms are then used to prioritize aid to the poorest regions and mobile subscribers. We evaluate this approach by studying Novissi, Togo's flagship anti-poverty program, which used these algorithms to determine eligibility for a rural assistance program that disbursed millions of dollars in COVID-19 relief aid.
Our analysis compares outcomes - including exclusion errors, total social welfare, and measures of fairness - under different targeting regimes. Relative to the other targeting options available to the Government of Togo at the time, the machine learning approach reduced errors of exclusion by between 2% and 50%. Relative to methods that require a comprehensive social registry (a hypothetical exercise; no such registry exists in Togo), the machine learning approach increases exclusion errors by 4-30%. These results highlight the potential for new data sources to contribute to humanitarian response efforts,
particularly in crisis settings when traditional data are missing or out of date.
This paper was distributed as Working Paper 29070, where an updated version may be available.
Bates, Du, and Wang show that a firm's ability to automate its workforce enhances operating flexibility, allowing for less conservative financial policies. Using an occupational measure of labor's susceptibility to automation, the researchers find that firms with a more substitutable workforce hold less cash, use more financial leverage, and pay higher dividends. Bates, Du, and Wang derive causal evidence exploiting the 2011-2012 Thailand hard drive crisis as a shock to the cost of automation. Following adverse shocks to cash flow from state tax increases, firms with an automatable workforce increase investment in equipment and software, reduce labor share in production, and experience a decline in operating leverage.