Search

The 2024 Index is our most comprehensive to date and arrives at an important moment when AI's influence on society has never been more pronounced. This year, we have broadened our scope to more extensively cover essential trends such as technical advancements in AI, public perceptions of the technology, and the geopolitical dynamics surrounding its development. Featuring more original data than ever before, this edition introduces new estimates on AI training costs, detailed analyses of the responsible AI landscape, and an entirely new chapter dedicated to AI's impact on science and medicine. The AI Index report tracks, collates, distills, and visualizes data related to artificial intelligence (AI). Our mission is to provide unbiased, rigorously vetted, broadly sourced data in order for policymakers, researchers, executives, journalists, and the general public to develop a more thorough and nuanced understanding of the complex field of AI. The AI Index is recognized globally as one of the most credible and authoritative sources for data and insights on artificial intelligence. Previous editions have been cited in major newspapers, including the The New York Times, Bloomberg, and The Guardian, have amassed hundreds of academic citations, and been referenced by high-level policymakers in the United States, the United Kingdom, and the European Union, among other places. This year's edition surpasses all previous ones in size, scale, and scope, reflecting the growing significance that AI is coming to hold in all of our lives.

Welcome to the sixth edition of the AI Index Report. This year, the report introduces more original data than any previous edition, including a new chapter on AI public opinion, a more thorough technical performance chapter, original analysis about large language and multimodal models, detailed trends in global AI legislation records, a study of the environmental impact of AI systems, and more. The AI Index Report tracks, collates, distills, and visualizes data related to artificial intelligence. Our mission is to provide unbiased, rigorously vetted, broadly sourced data in order for policymakers, researchers, executives, journalists, and the general public to develop a more thorough and nuanced understanding of the complex field of AI. The report aims to be the world's most credible and authoritative source for data and insights about AI.

In September 2016, Stanford's "One Hundred Year Study on Artificial Intelligence" project (AI100) issued the first report of its planned long-term periodic assessment of artificial intelligence (AI) and its impact on society. It was written by a panel of 17 study authors, each of whom is deeply rooted in AI research, chaired by Peter Stone of the University of Texas at Austin. The report, entitled "Artificial Intelligence and Life in 2030," examines eight domains of typical urban settings on which AI is likely to have impact over the coming years: transportation, home and service robots, healthcare, education, public safety and security, low-resource communities, employment and workplace, and entertainment. It aims to provide the general public with a scientifically and technologically accurate portrayal of the current state of AI and its potential and to help guide decisions in industry and governments, as well as to inform research and development in the field. The charge for this report was given to the panel by the AI100 Standing Committee, chaired by Barbara Grosz of Harvard University., Comment: 52 pages, https://ai100.stanford.edu/2016-report

Welcome to the fifth edition of the AI Index Report! The latest edition includes data from a broad set of academic, private, and nonprofit organizations as well as more self-collected data and original analysis than any previous editions, including an expanded technical performance chapter, a new survey of robotics researchers around the world, data on global AI legislation records in 25 countries, and a new chapter with an in-depth analysis of technical AI ethics metrics. The AI Index Report tracks, collates, distills, and visualizes data related to artificial intelligence. Its mission is to provide unbiased, rigorously vetted, and globally sourced data for policymakers, researchers, executives, journalists, and the general public to develop a more thorough and nuanced understanding of the complex field of AI. The report aims to be the world's most credible and authoritative source for data and insights about AI.

In 1950, Alan Turing proposed an imitation game as the ultimate test of whether a machine was intelligent: could a machine imitate a human so well that its answers to questions indistinguishable from a human. Ever since, creating intelligence that matches human intelligence has implicitly or explicitly been the goal of thousands of researchers, engineers, and entrepreneurs. The benefits of human-like artificial intelligence (HLAI) include soaring productivity, increased leisure, and perhaps most profoundly, a better understanding of our own minds. But not all types of AI are human-like. In fact, many of the most powerful systems are very different from humans. So an excessive focus on developing and deploying HLAI can lead us into a trap. As machines become better substitutes for human labor, workers lose economic and political bargaining power and become increasingly dependent on those who control the technology. In contrast, when AI is focused on augmenting humans rather than mimicking them, then humans retain the power to insist on a share of the value created. Furthermore, augmentation creates new capabilities and new products and services, ultimately generating far more value than merely human-like AI. While both types of AI can be enormously beneficial, there are currently excess incentives for automation rather than augmentation among technologists, business executives, and policymakers., Comment: Forthcoming in Daedalus, April 2022. Posted with permission

We suggest a methodology that allows statistical agencies to form approximations to the benefits that flow to households from new free goods. The present production-oriented GDP measures are not satisfactory for measuring real household consumption and will be increasingly inaccurate as free goods, such as those made possible by the digital revolution, become more important. Advertising expenditures are not an adequate substitute for measuring the benefits of new goods to the household sector. Instead, we need to draw on estimates such as those provided by choice experiments.

Partnering with the US Census Bureau, we implement a new survey of "structured" management practices in two waves of 35,000 manufacturing plants in 2010 and 2015. We fnd an enormous dispersion of management practices across plants, with 40 percent of this variation across plants within the same frm. Management practices account for more than 20 percent of the variation in productivity, a similar, or greater, percentage as that accounted for by R&D, ICT, or human capital. We fnd evidence of two key drivers to improve management. The business environment, as measured by right-to-work laws, boosts incentive management practices. Learning spillovers, as measured by the arrival of large "Million Dollar Plants" in the county, increases the management scores of incumbents.

Anecdotes abound suggesting that the use of predictive analytics boosts firm performance. However, large-scale representative data on this phenomenon have been lacking. Working with the Census Bureau, we surveyed over 30,000 American manufacturing establishments on their use of predictive analytics and detailed workplace characteristics. We find that productivity is significantly higher among plants that use predictive analytics—up to $918,000 higher sales compared to similar competitors. Furthermore, both instrumental variables estimates and the timing of gains suggest a causal relationship. However, we find that the productivity pay-off only occurs when predictive analytics are combined with at least one of three workplace complements: significant accumulation of IT capital, educated workers, or workplaces designed for high flow-efficiency production. Our findings support claims that predictive analytics can substantially boost performance, while also explaining why some firms see no benefits at all.

We suggest a methodology that allows statistical agencies to form approximations to the benefits that flow to households from new free goods. The present production-oriented GDP measures are not satisfactory for measuring real household consumption and will be increasingly inaccurate as free goods, such as those made possible by the digital revolution, become more important. Advertising expenditures are not an adequate substitute for measuring the benefits of new goods to the household sector. Instead, we need to draw on estimates such as those provided by choice experiments.

© 2019 National Academy of Sciences. All Rights Reserved. Rapid advances in artificial intelligence (AI) and automation technologies have the potential to significantly disrupt labor markets. While AI and automation can augment the productivity of some workers, they can replace the work done by others and will likely transform almost all occupations at least to some degree. Rising automation is happening in a period of growing economic inequality, raising fears of mass technological unemployment and a renewed call for policy efforts to address the consequences of technological change. In this paper we discuss the barriers that inhibit scientists from measuring the effects of AI and automation on the future of work. These barriers include the lack of high-quality data about the nature of work (e.g., the dynamic requirements of occupations), lack of empirically informed models of key microlevel processes (e.g., skill substitution and human–machine complementarity), and insufficient understanding of how cognitive technologies interact with broader economic dynamics and institutional mechanisms (e.g., urban migration and international trade policy). Overcoming these barriers requires improvements in the longitudinal and spatial resolution of data, as well as refinements to data on workplace skills. These improvements will enable multidisciplinary research to quantitatively monitor and predict the complex evolution of work in tandem with technological progress. Finally, given the fundamental uncertainty in predicting technological change, we recommend developing a decision framework that focuses on resilience to unexpected scenarios in addition to general equilibrium behavior.

© 2020 Copyright held by the owner/author(s). The recent wave of AI and automation has been argued to differ from previous General Purpose Technologies (GPTs), in that it may lead to rapid change in occupations' underlying task requirements and persistent technological unemployment. In this paper, we apply a novel methodology of dynamic task shares to a large dataset of online job postings to explore how exactly occupational task demands have changed over the past decade of AI innovation, especially across high, mid and low wage occupations. Notably, big data and AI have risen significantly among high wage occupations since 2012 and 2016, respectively. We built an ARIMA model to predict future occupational task demands and showcase several relevant examples in Healthcare, Administration, and IT. Such task demands predictions across occupations will play a pivotal role in retraining the workforce of the future.

AI is undergoing a paradigm shift with the rise of models (e.g., BERT, DALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a wide range of downstream tasks. We call these models foundation models to underscore their critically central yet incomplete character. This report provides a thorough account of the opportunities and risks of foundation models, ranging from their capabilities (e.g., language, vision, robotics, reasoning, human interaction) and technical principles(e.g., model architectures, training procedures, data, systems, security, evaluation, theory) to their applications (e.g., law, healthcare, education) and societal impact (e.g., inequity, misuse, economic and environmental impact, legal and ethical considerations). Though foundation models are based on standard deep learning and transfer learning, their scale results in new emergent capabilities,and their effectiveness across so many tasks incentivizes homogenization. Homogenization provides powerful leverage but demands caution, as the defects of the foundation model are inherited by all the adapted models downstream. Despite the impending widespread deployment of foundation models, we currently lack a clear understanding of how they work, when they fail, and what they are even capable of due to their emergent properties. To tackle these questions, we believe much of the critical research on foundation models will require deep interdisciplinary collaboration commensurate with their fundamentally sociotechnical nature., Comment: Authored by the Center for Research on Foundation Models (CRFM) at the Stanford Institute for Human-Centered Artificial Intelligence (HAI). Report page with citation guidelines: https://crfm.stanford.edu/report.html

Welcome to the fourth edition of the AI Index Report. This year we significantly expanded the amount of data available in the report, worked with a broader set of external organizations to calibrate our data, and deepened our connections with the Stanford Institute for Human-Centered Artificial Intelligence (HAI). The AI Index Report tracks, collates, distills, and visualizes data related to artificial intelligence. Its mission is to provide unbiased, rigorously vetted, and globally sourced data for policymakers, researchers, executives, journalists, and the general public to develop intuitions about the complex field of AI. The report aims to be the most credible and authoritative source for data and insights about AI in the world.

Most of the empirical evidence on social advertising effectiveness focuses on a single product at a time. As a result, little is known about how the effectiveness of social advertising varies across product categories or product characteristics. We therefore collaborated with a large online social network to conduct a randomized field experiment measuring social ad effectiveness across 71 products in 25 categories among more than 37 million users. We found some product categories, like clothing, cars, and food, exhibited significantly stronger social advertising effectiveness than other categories, like financial services, electrical appliances, and mobile games. More generally, we found that status goods, which rely on status-driven consumption, displayed strong social advertising effectiveness. Meanwhile, social ads for experience goods, which rely on informational social influence, did not perform any better or worse than social ads for their theoretical counterparts, search goods. Social advertising effectiveness also significantly varied across the relative characteristics of ad viewers and their friends shown in ads. Understanding the heterogeneous effects of social advertising across products can help marketers differentiate their social advertising strategies and lead researchers to more nuanced theories of social influence in product evaluation.

Artificial intelligence (AI) is surpassing human performance in a growing number of domains. However, there is limited evidence of its economic effects. Using data from a digital platform, we study a key application of AI: machine translation. We find that the introduction of a new machine translation system has significantly increased international trade on this platform, increasing exports by 10.9%. Furthermore, heterogeneous treatment effects are consistent with a substantial reduction in translation costs. Our results provide causal evidence that language barriers significantly hinder trade and that AI has already begun to improve economic efficiency in at least one domain.

An overly simplistic reliance on the utility model risks blinding us to the real opportunities and challenges of cloud computing.

The recent wave of AI and automation has been argued to differ from previous General Purpose Technologies (GPTs), in that it may lead to rapid change in occupations' underlying task requirements and persistent technological unemployment. In this paper, we apply a novel methodology of dynamic task shares to a large dataset of online job postings to explore how exactly occupational task demands have changed over the past decade of AI innovation, especially across high, mid and low wage occupations. Notably, big data and AI have risen significantly among high wage occupations since 2012 and 2016, respectively. We built an ARIMA model to predict future occupational task demands and showcase several relevant examples in Healthcare, Administration, and IT. Such task demands predictions across occupations will play a pivotal role in retraining the workforce of the future., Comment: 9 pages, 5 figures, 6 tables, Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES), 2020

Gross domestic product (GDP) and derived metrics such as productivity have been central to our understanding of economic progress and well-being. In principle, changes in consumer surplus provide a superior, and more direct, measure of changes in wellbeing, especially for digital goods. In practice, these alternatives have been difficult to quantify. We explore the potential of massive online choice experiments to measure consumer surplus. We illustrate this technique via several empirical examples which quantify the valuations of popular digital goods and categories. Our examples include incentive-compatible discrete-choice experiments where online and laboratory participants receive monetary compensation if and only if they forgo goods for predefined periods. For example, the median user needed a compensation of about $48 to forgo Facebook for 1 mo. Our overall analyses reveal that digital goods have created large gains in well-being that are not reflected in conventional measures of GDP and productivity. By periodically querying a large, representative sample of goods and services, including those which are not priced in existing markets, changes in consumer surplus and other new measures of well-being derived from these online choice experiments have the potential for providing cost-effective supplements to the existing national income and product accounts.

GDP is a measure of production, and yet it is widely used as a proxy for well-being. It is particularly ill-suited for assessing the contributions of digital goods which are free to consumers and thus excluded from GDP measures. This underscores the need to develop new measures of well-being which can assess not only the contributions of digital goods but also welfare more generally. In Brynjolfsson, Eggers, and Gannamaneni (2017), we propose a new way of measuring consumer welfare using massive online choice experiments. This brief paper motivates the need for such an approach and introduces the method.

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 provide a systematic empirical study of the diffusion and adoption patterns of data-driven decision making (DDD) in the U.S. Using data collected by the Census Bureau for a large representative sample of manufacturing plants, we find that DDD rates nearly tripled (11%-30%) between 2005 and 2010. This rapid diffusion, along with results from a companion paper, are consistent with case-based evidence that DDD tends to be productivity-enhancing. Yet certain plants are significantly more likely to adopt than others. Key correlates of adoption are size, presence of potential complements such as information technology and educated workers, and firm learning.

We econometrically evaluate information worker productivity at a midsize executive recruiting firm and assess whether the knowledge that workers accessed through their electronic communication networks enabled them to multitask more productively. We estimate dynamic panel data models of multitasking, knowledge networks, and productivity using several types of micro-level data: (a) direct observation of more than 125,000 email messages over a period of 10 months; (b) detailed accounting data on individuals' project output and team membership for more than 1,300 projects spanning five years; and (c) survey and interview data about the same workers' IT skills, IT use, and information sharing. We find that (1) more multitasking is associated with more project output, but diminishing marginal returns, and (2) recruiters whose network contacts have heterogeneous knowledge—an even distribution of expertise over many project types—are less productive on average but more productive when juggling diverse multitasking portfolios. These results show how multitasking affects productivity and how knowledge networks, enabled by IT, can improve worker performance. The methods developed can be replicated in other settings, opening new frontiers for research on social networks and IT value.

We gather detailed data on organizational practices and information technology (IT) use at 253 firms to examine the hypothesis that external focus—the ability of a firm to detect and therefore respond to changes in its external operating environment—increases returns to IT, especially when combined with decentralized decision making. First, using survey-based measures, we find that external focus is correlated with both organizational decentralization, and IT investment. Second, we find that a cluster of practices including external focus, decentralization, and IT is associated with improved product innovation capabilities. Third, we develop and test a three-way complementarities model that indicates that the combination of external focus, decentralization, and IT is associated with significantly higher productivity in our sample. We also introduce a new set of instrumental variables representing barriers to IT-related organizational change and find that our results are robust when we account for the potential endogeneity of organizational investments. Our results may help explain why firms that operate in information-rich environments such as high-technology clusters or areas with high worker mobility have experienced especially high returns to IT investment and suggest a set of practices that some managers may be able to use to increase their returns from IT investments., Massachusetts Institute of Technology. Center for Digital Business, National Science Foundation (U.S.) (Grant IRI-9733877)

The Internet and related information technologies are transforming the distribution of product sales across products, and these effects are likely to grow in coming years. Both the Long Tail and the Superstar effect are manifestations of these changes, yet researchers lack consistent metrics or models for integrating and extending their insights and predictions. In this paper, we begin with a taxonomy of the technological and nontechnological drivers of both Long Tails and Superstars and then define and compare the key metrics for analyzing these phenomena. The core of the paper describes a large and promising set of questions forming a research agenda. Important opportunities exist for understanding future changes in sales concentration patterns; the impact on supply chains (including cross-channel competition, competition within the Internet channel, implications for the growth of firms, and the balance of power within the supply chain); implications for pricing, promotion, and product design; and, ultimately, the potential effects on society in general. Our approach provides an introduction to some of the relevant research findings and allows us to identify opportunities for cross-pollination of methods and insights from related research topics., MIT Center for Digital Business

We test for three-way complementarities among information technology (IT), performance pay, and human resource (HR) analytics practices. We develop a principal–agent model examining how these practices work together as an incentive system that produces a larger productivity premium when the practices are implemented in concert rather than separately. We assess our model by combining fine-grained data on human capital management (HCM) software adoption over 11 years with detailed survey data on incentive systems and HR analytics practices for 189 firms. We find that the adoption of HCM software is greatest in firms that have also adopted performance pay and HR analytics practices. Furthermore, HCM adoption is associated with a large productivity premium when it is implemented as a system of organizational incentives, but has less benefit when adopted in isolation. The system of three-way complements produces disproportionately greater benefits than pairwise interactions, highlighting the importance of including all three complements. Productivity increases significantly when the HCM systems “go live” but not when they are purchased, which can be years earlier. This helps rule out reverse causality as an explanation for our findings., MIT Center for Digital Business

Many markets have historically been dominated by a small number of best-selling products. The Pareto principle, also known as the 80/20 rule, describes this common pattern of sales concentration. However, information technology in general and Internet markets in particular have the potential to substantially increase the collective share of niche products, thereby creating a longer tail in the distribution of sales. This paper investigates the Internet's “long tail” phenomenon. By analyzing data collected from a multichannel retailer, it provides empirical evidence that the Internet channel exhibits a significantly less concentrated sales distribution when compared with traditional channels. Previous explanations for this result have focused on differences in product availability between channels. However, we demonstrate that the result survives even when the Internet and traditional channels share exactly the same product availability and prices. Instead, we find that consumers' usage of Internet search and discovery tools, such as recommendation engines, are associated with an increase the share of niche products. We conclude that the Internet's long tail is not solely due to the increase in product selection but may also partly reflect lower search costs on the Internet. If the relationships we uncover persist, the underlying trends in technology portend an ongoing shift in the distribution of product sales., MIT Center for Digital Business, National Science Foundation (U.S.) (Grant IIS-0085725)

This version published in "Innovation Incentives for Information Goods," Innovation Policy and the Economy eds. Adam B. Jaffe, Josh Lerner and Scott Stern, 2007 vol. 7. Previous version in published as “Slicing the Gordian Knot: A Novel Mechanism for Providing Innovation Incentives for Digital Goods” in Proceedings of the International Conference on Information Systems, December, 2004., Innovations can often be targeted to be more valuable for some consumers than others. This is especially true for digital information goods. We show that the traditional price system not only results in significant deadweight loss, but also provides incorrect incentives to the creators of these innovations. In contrast, we propose and analyze a profit-maximizing mechanism for bundles of digital goods which is more efficient and more accurately provides innovation incentives for information goods. Our “statistical couponing mechanism” does not rely on the universal excludability of information goods, which creates substantial deadweight loss, but instead estimates social value created from new goods and innovations by offering coupons to a relatively small sample of representative consumers. We find that the statistical couponing mechanism can operate with less than 0.1% of the deadweight loss of the traditional price-based system, while more accurately aligning incentives with social value., The MIT Center for Digital Business, the National Bureau of Economic Research and that National Science Foundation (IIS-0085725) provided generous research funding.

We examine the drivers of diffusion of information through organizations and the effects on performance. In particular, we ask: What predicts the likelihood of an individual becoming aware of a strategic piece of information, or becoming aware of it sooner? Do different types of information exhibit different diffusion patterns, and do different characteristics of social structure, relationships and individuals in turn affect access to different kinds of information? Does better access to information predict an individual’s ability to complete projects or generate revenue? We hypothesize that the dual effects of content and structure jointly predict the diffusion path of information, and ultimately performance. To test our hypotheses, we characterize the social network of a medium sized executive recruiting firm using accounting data on project co-work relationships and ten months of email traffic observed over two five month periods. We identify two distinct types of information diffusing over this network – ‘event news’ and ‘discussion topics’ – by their usage characteristics, and observe several thousand diffusion processes of each type of information from their original first use to their varied recipients over time. We then test the effects of network structure and functional and demographic characteristics of dyadic relationships on the likelihood of receiving each type of information and receiving it more quickly. Our results demonstrate that the diffusion of news, characterized by a spike in communication and rapid, pervasive diffusion through the organization, is influenced by demographic and network factors but not by functional relationships (e.g. prior co-work, authority) or the strength of ties. In contrast, diffusion of discussion topics, which exhibit more shallow diffusion characterized by ‘back-and-forth’ conversation, is heavily influenced by functional relationships and the strength of ties, as well as demographic and network factors. Discussion to, The National Science Foundation, Cisco Systems, France Telecom and the MIT Center for Digital Business

"First Draft: May, 1993, Revised, January 1994."

Title from cover., "October 2001."

"November 1989."

"First Draft: May, 1993, Revised, January 1994."

"November 1989."

Title from cover., "October 2001."

Thesis (Ph. D.)--Massachusetts Institute of Technology, Sloan School of Management, 1991., Includes bibliographical references., by Erik Brynjolfsson., Ph.D.

by Erik Brynjolfsson., Thesis (Ph. D.)--Massachusetts Institute of Technology, Sloan School of Management, 1991., Includes bibliographical references.

Price dispersion among commodity goods is typically attributed to consumer search costs. We explore the magnitude of consumer search costs using a data set obtained from a major Internet shopbot. For the median consumer, the benefits to searching lower screens are $2.24 while the cost of an exhaustive search of the offers is a maximum of $2.03. Interestingly, in our setting, consumers who search more intensively are less price sensitive than other consumers, reflecting their increased weight on retailer differentiation in delivery time and reliability. Our results demonstrate that even in this nearly-perfect market, substantial price dispersion can exist in equilibrium from consumers preferences over both price and non-price attributes

We explore the effect of computerization on productivity and output growth using data from 527 large US firms over 1987-1994. We find that computerization makes a contribution to measured productivity and output growth in the short term (using one year differences) that is consistent with normal returns to computer investments. However, the productivity and output contributions associated with computerization are up to five times greater over long periods (using five to seven year differences). The results suggest that the observed contribution of computerization is accompanied by relatively large and time-consuming investments in complementary inputs, such as organizational capital, that may be omitted in conventional calculations of productivity. The large long-run contribution of computers and their associated complements that we uncover may partially explain the subsequent investment surge in computers in the late 1990s.

Price dispersion among commodity goods is typically attributed to consumer search costs. We explore the magnitude of consumer search costs using a data set obtained from a major Internet shopbot. For the median consumer, the benefits to searching lower screens are $2.24 while the cost of an exhaustive search of the offers is a maximum of $2.03. Interestingly, in our setting, consumers who search more intensively are less price sensitive than other consumers, reflecting their increased weight on retailer differentiation in delivery time and reliability. Our results demonstrate that even in this nearly-perfect market, substantial price dispersion can exist in equilibrium from consumers preferences over both price and non-price attributes

Price dispersion among commodity goods is typically attributed to consumer search costs. We explore the magnitude of consumer search costs using a data set obtained from a major Internet shopbot. For the median consumer, the benefits to searching lower screens are $2.24 while the cost of an exhaustive search of the offers is a maximum of $2.03. Interestingly, in our setting, consumers who search more intensively are less price sensitive than other consumers, reflecting their increased weight on retailer differentiation in delivery time and reliability. Our results demonstrate that even in this nearly-perfect market, substantial price dispersion can exist in equilibrium from consumers preferences over both price and non-price attributes

We explore the effect of computerization on productivity and output growth using data from 527 large US firms over 1987-1994. We find that computerization makes a contribution to measured productivity and output growth in the short term (using one year differences) that is consistent with normal returns to computer investments. However, the productivity and output contributions associated with computerization are up to five times greater over long periods (using five to seven year differences). The results suggest that the observed contribution of computerization is accompanied by relatively large and time-consuming investments in complementary inputs, such as organizational capital, that may be omitted in conventional calculations of productivity. The large long-run contribution of computers and their associated complements that we uncover may partially explain the subsequent investment surge in computers in the late 1990s.