Technological anxiety is at least as old as the industrial revolution, so the rapid development of generative artificial intelligence (genAI) products has spurred research and analysis on the impact this technology will have on labor markets. In chapters in a new book, researchers examine how the structure of tasks can facilitate or impede the adoption of genAI, how workers of different types choose to use genAI, and where workers are likely to look for jobs if they are displaced from their work due to genAI. GenAI will likely widen the "cone of automation" by substituting for labor in more complex work or in work that occurs less frequently, the authors conclude.
The chapters, written by researchers at Carnegie Mellon University, the University of Southern California, and the University of Pennsylvania, appear in The Oxford Handbook of the Foundations and Regulation of Generative AI.
"Our conceptualization of a cone of automation provides a simple visual representation automation is expected to occur, given the characteristic of a technology," explains Ramayya Krishnan, professor of management science and information systems at and emeritus dean of Carnegie Mellon's Heinz College, who coauthored the chapter. "Relevant dimensions are the overall output, or frequency at which a step needs to be completed, and the length of the step as currently configured in production."
The cone of automation highlights several facts: 1) Automation is more likely to occur in steps performed at a high frequency; this is intuitive since the benefits of a machine are more likely to be realized when the machine is working at high capacity. 2) Automation is more likely to occur for "middle-length" steps; only when output grows does it become more likely to automate easy steps. 3) People are more likely to be an economic advantage when dealing with particularly long, complex steps.
GenAI will likely widen the cone of automation by substituting for labor in more complex work or in work that occurs less frequently, the authors suggest. When the costs of failure are high, businesses will probably adopt less genAI due to its randomness. In this case, the cone of automation would narrow and genAI would play an explicitly complementary role that involves having a person oversee its work.
"GenAI differs considerably from classical machines in that it is more general and more useful but also more prone to errors," notes Laurence Ales, professor of economics at Carnegie Mellon's Tepper School of Business, who coauthored the chapter. "These features inform the potential patterns businesses will use in adopting genAI, including whether it will substitute for or complement existing workers."
The technical feasibility of automation using genAI is not enough to understand these patterns, according to the authors. The economic conditions for adoption depend on the interaction of technical features with process structure. The cost and benefit of dividing tasks drive how firms currently organize work and define jobs, and measures of occupational exposure to genAI or other technologies must consider the relative frequency and separability of tasks.
In the long run, the use of genAI will influence the quality of data available for training future models, the authors note. GenAI is often mediated by human users with different levels of skill: The more genAI is used by workers with less ability to identify and correct errors in output (while increasing the quantity of this low-quality output), the more the quality of future training corpuses is likely to degrade, they predict.
"We may expect a divergence in genAI quality, in which lower data quality further reduces the complementarity of the technology with high skill, whereas contexts with high error standards will see narrower and perhaps slower diffusion of genAI but higher long-run complementarity with high skill and high data quality," says Christophe Combemale, research professor of engineering and public policy at Carnegie Mellon, who coauthored the chapter.
Finally, the authors consider the potential shape of occupational disruption due to genAI. A network view of occupations is needed to anticipate outcomes for disrupted workers, they suggest. Even occupations not directly disrupted by genAI may experience competition and wage losses if they become targets for workforce transitions out of disrupted occupations. Conversely, the resilience of labor markets in providing employment for disrupted workers will depend on having a sufficient density of alternative, less AI-substitutable occupations into which workers can transition.
