Experts: Erik Brynolfsson (Stanford), Lee Branstetter (CMU), and Dashun Wang (Northwestern)
Brynjolfsson, along with Li, Bana and Steffen are exploring the development, integration, and impact of AI on the transformation of the economy and work. They propose two complementary research projects that will greatly improve understanding of the economic effects and potential of AI, helping leaders make informed, data-driven decisions that can lead to more equitable outcomes. First, Brynjolfsson and team propose a detailed examination of AI adoption by firms in America, including a focus on AI adoption by firms involved in the three advanced manufacturing deep dives: biotech, energy, semiconductors. The Brynjolfsson team is analyzing data from the U.S. Census Bureau’s 2018 Annual Business Survey of over 850,000 firms to establish a number of stylized facts about early AI adoption in the U.S. The preliminary insights of Brynjolfsson and team show that while less than 6 percent of firms use any of the AI technologies that they measure, adoption is prevalent in a subset of distinctive firms (Zolas et al., 2020). This work is essential to understanding both the firm-level and macroeconomic impacts of AI adoption in the US; such metrics will be highly relevant in examining the impacts of AI investments and other government programs and policies. Second, AI’s impact on workers and jobs is a related and important complement to firm-level analysis. Job postings provide unique insights into the demand for skills, tasks, and occupations. Using the full text of data from over 200 million online job postings, Brynjolfsson and team can train and evaluate a natural language processing (NLP) model, which will allow further insights into how jobs are changing with respect to the adoption of emerging technologies (especially AI). This work could help workers, employers, policymakers, and industry leaders better prepare for and leverage new technologies.
Branstetter and Hovy will leverage frontier natural language processing and machine learning tools to map the evolution of innovative capacity in critical technologies across geography, time, and technological domains. This project will focus initially on AI, but the methods can be applied to other critical and emerging technologies. Branstetter and Hovy will begin by using machine learning and natural language processing techniques to identify AI-related U.S. patents as a measure of AI invention. They plan to utilize their collaboration with the U.S. Census Bureau to measure the impact of AI inventions on inventing firms, including those that are not publicly traded. By tracking the movement of prolific AI inventors across firm and organizational boundaries, they can measure the diffusion of AI inventive capabilities across firms, industries, and geography. The next phase of this project will exploit that firms undertaking a serious investment in internal AI capabilities may not always file a patent, but they will often have to hire a critical mass of PhD-level experts. Branstetter and Hovy plan to mine publication data to identify PhD students, categorize their research specialties, and measure the quality of their academic programs and faculty advisers. The collaboration of Branstetter and Hovy with the U.S. Census Bureau allows them to measure the impact of the emergence within a firm of a critical mass of PhD-level AI experts on firm-level output, employment, and productivity.
The Stanford and CMU teams will learn across each other, leveraging the different approaches to data (survey versus machine learning and natural language processing) and algorithms to enhance the possibility frontier in the timely and dynamic identification of capabilities in critical technologies and their impact on productivity and jobs.
Wang builds on this work by connecting the deep dive in AI to understand how AI impacts the rate and direction of scientific discovery and commercialization. The development of critical technology often begins from upstream scientific developments. Here, for the first year, Wang will develop a general framework to assess the overall benefits of AI across scientific disciplines, with a special focus on the three other advanced manufacturing deep dives (biotechnology, energy storage, and semiconductors.) Given that we know relatively little about the impact of AI across scientific disciplines, and the commercialization of that science, Wang and team plan to build a quantitative and systematic framework for estimating the impact of AI in driving scientific discoveries and their commercialization and to understand potential heterogeneities across scientific disciplines. They will apply NLP techniques to analyze two large-scale datasets, sourced respectively from the Microsoft Academic Graph (MAG), covering about 88 million papers, and 7.6 million patents granted by the U.S. Patent and Trademark Office (USPTO). They will also integrate their framework with two additional large-scale datasets capturing 5.2 million university course syllabi, and 5.4 million research grants to further examine potential alignment between the impact of AI and education and funding investments of AI. Wang and team plan to calculate the direct use of AI in each field (i.e., direct impact) by developing an “n-gram” based framework. They plan to estimate the potential impact of AI by developing an “AI capability-field task” framework. They will infer AI capabilities (i.e., what AI can do) and basic tasks for a field (i.e., what the field does) by extracting verb-noun pairs (e.g., “detect pattern”) from AI-related and field papers/patents using NLP techniques, respectively. Wang and team will estimate a field’s potential impact of AI by calculating the overlap between its current tasks and cumulative AI capabilities. They then plan to explore whether the estimated impact of AI on science and its commercialization is commensurate with the funding of AI in research fields and the education of AI in classrooms. Linking to the equity team, they also plan to explore whether AI impacts in science are accruing inequitably for different demographic groups and assess potential gender and racial gaps in benefiting from AI advances.