People have yet to reach a consensus on any AI-related issues, but one expectation seems to have gained traction among most companies, investors, and analysts: AI will boost productivity across many sectors. Even if these gains can be realized, is it worth it?
Many observers, including myself, expect that AI will significantly enhance productivity, and the increasing number of cases providing preliminary evidence seems to confirm this. Given the rapid expansion of AI capabilities, the decreasing costs of training and using large models, and the growing availability of open-source tools and systems, AI appears to be effectively applicable to nearly every industry and profession.
However, the effective use of AI is not guaranteed, nor will it happen overnight, as it is influenced by factors such as acquisition, dissemination, and the learning curve. Even after overcoming these barriers, it is far from certain that the productivity boosts from AI will lead to widespread benefits in terms of employment and income. This will depend on the development of both AI tools and the labor market.
AI tools are expanding rapidly. If most of the new capabilities focus on replicating human abilities to replace workers, then the productivity gains could lead to negative distributional effects. As noted in a recent paper by two digital economy researchers from Stanford University and MIT, current machine learning benchmarks are overly focused on automation and seldom consider humans in the evaluation process. To prevent AI from becoming a “parlor game,” they suggest the developer community adopt a “centaur-style evaluation,” where humans and AI systems collaborate to solve tasks. This would shift the development direction of machine learning towards enhancing functionality or strengthening human-machine collaboration, rather than just automation.
To ensure that the majority of the public can benefit from AI, we must focus on the labor market. For example, in the U.S., about 20% of the workforce is employed in tradable sectors, while the remaining 80% works in non-tradable service sectors like government, education, hospitality, traditional retail, and construction.
Over the past 30 years, the productivity and income gap between tradable and non-tradable sectors in the U.S. has been widening. Overall, the tradable sector has seen faster productivity growth and higher income levels. As manufacturing employment has stabilized, the value added to the economy has continued to grow. Without careful attention, AI could widen the gap between tradable and non-tradable sectors, leading to a sharp increase in inequality. Only when AI can be effectively applied to mid- and low-income jobs in both tradable and non-tradable sectors can overall economic productivity be increased and income growth be broadly distributed. Therefore, efforts must be made to guide AI towards enhancing collaboration across industries and income levels.
Some positive signals are emerging in this area. The U.S. Defense Advanced Research Projects Agency (DARPA) has held competitions focused on human-machine collaboration, such as using robots to enhance human physical capabilities and navigating robots in complex and rapidly changing physical environments. More work needs to be done in the future, including funding AI foundational research by various stakeholders, including the government. These efforts should focus on enhancing functionality and strengthening human-machine collaboration, while introducing relevant incentives for developers.
Ensuring that the AI toolkit benefits various industries and income groups by enhancing or fostering collaboration must be a top priority. But this alone cannot guarantee that AI will bring widespread prosperity, because general equilibrium effects are still at play. In the previous digitalization wave, many routine, codifiable jobs were automated. Additionally, globalization led to the outsourcing of manufacturing production, resulting in many middle-class workers being displaced, with few opportunities to transition into higher-productivity and higher-income jobs.
In the upcoming AI transformation period, productivity increases will lead to lower costs, and combined with normal competitive pressure, prices will drop. If demand elasticity in a certain industry is less than 1, job opportunities will decrease. Of course, other industries with higher demand elasticity will see an increase in employment. However, the movement of workers between industries signifies significant upheaval, and there may be a short-term risk of labor supply exceeding demand, which will erode workers’ bargaining power.
As many have pointed out, providing income and skills transition support for these groups is critical, and AI-driven tools are likely to offer assistance in retraining and acquiring new skills. Meanwhile, policymakers should create labor demand in a manner similar to the post-Great Depression era in the U.S.
For the U.S., this presents a win-win opportunity. For various reasons, the U.S. has fallen behind other countries in infrastructure development and upgrades. Addressing this situation will create numerous job opportunities and labor demand, providing a buffer for the upcoming AI-driven transformation.