[Christophe Andre] Will AI hit employment, raise productivity, and increase inequality?

The release of ChatGPT in November 2022 has raised great hopes that artificial intelligence (AI) can contribute to solving problems in many fields and lift productivity, but also fears that many jobs may disappear, and that income inequality could rise further. AI is commonly seen as a general-purpose technology, like the internal combustion engine, early electricity-based technologies, and computers. Such technologies have the potential to disrupt large parts of the economy, displacing many workers. Every wave of innovation in history has generated fears of “technological unemployment” but so far economies have generally succeeded in creating more jobs than those that were destroyed, even though transition phases have often been painful.

Automation generally leads to productivity gains, which generate income that can be spent on new products and services, which in turn creates demand for labor. Employment in difficult to automate activities rises with aggregate demand and new occupations are being created. David Autor, from the Massachusetts Institute of Technology, and co-authors, find that about 60 percent of 2018 US jobs did not exist in 1940. Reallocation of labpr associated with technological change challenges workers, who often need reskilling and upskilling. Structural change may also be associated with macroeconomic and financial instability if economies struggle to adapt. The income distribution is affected, as illustrated by the polarization of labor markets in many advanced economies over the past decades.

Will labor market developments in the AI era follow historical patterns, or will AI be a game changer? Like many digital tools before it, AI offers large potential for deployment across a wide range of economic activities. Until recently, mainly routine tasks could be automated. AI offers possibilities to automate non-routine cognitive tasks, thanks to its ability to learn by itself from huge datasets and integrate tacit knowledge. Some researchers believe that machines could ultimately outperform humans in almost all activities. Nevertheless, even if one believes such prophecies, they are probably far away in the future. At the current juncture, AI can automate some cognitive tasks and complement humans in performing more complex ones.

Several studies have investigated different occupations’ exposure to AI, by linking some AI applications to skills used in specific professions. This is a good starting point for thinking about the potential impact of AI on jobs, productivity, and inequality. Research from the OECD Global Forum on Productivity suggests that jobs in knowledge-intensive services, such as finance, advertising, consulting, and information and communication are the most likely to be affected by AI. Conversely, workers in industries like mining and construction, or relatively low knowledge-intensity services, such as administration and support, transportation, and water and waste, have limited exposure to AI. This has prompted fears that technological development, after automating routine tasks, which affected mainly blue-collar workers, would go on to automate even knowledge-intensive tasks, threatening the jobs of high-skilled workers.

A crucial question is whether AI can mainly automate tasks or whether it also has potential to augment workers’ abilities to perform their jobs. In other words, whether AI is mostly a substitute or a complement to human labor. According to a World Economic Forum report, both aspects are important, and their strength varies across economic sectors. Tasks are considered automatable if they do not require creatively solving ambiguous problems, working with others in real time or validating outputs. The potential for augmentation of worker capabilities is particularly high in knowledge-intensity services. For example, in information and technology services, nearly 32 percent of the time spent on all tasks could be subject to automation, but the potential for augmentation is even higher, reaching almost 37 percent of the time spent on all tasks. In media and publishing, the potential for augmentation dominates that of automation even more, at 32 percent versus 24 percent of the time spent on all tasks. Hence, while AI is likely to displace some workers, it will also boost the productivity of others. To give more concrete examples, while AI can improve medical diagnosis and help build legal cases, it is unlikely to replace physicians and lawyers anytime soon. AI provides skilled professionals with new valuable tools, but in many circumstances, human intervention is still necessary, especially in high-stakes decisions, not least because AI is sometimes subject to hallucinations, producing inaccurate or misleading results.

Recent studies point to significant improvements in workers’ performance due to AI in some specific tasks, ranging from 14 percent in customer services to 56 percent in coding. This suggests that if AI were to be widely adopted across the economy, productivity may rise substantially. However, translating micro-level performance into economy-wide productivity is challenging. Estimates of AI’s potential labor productivity boost over the next decade in advanced economies generally vary from about 0.1 to more than 2.5 percentage points per year. To put these numbers into perspective, US nonfarm labor productivity grew at an average annual rate of 2.1 percent between 1947 and 2023, and 1.5 percent since 2006. Hence, according to the most optimistic estimates, AI would lift productivity growth well-above historical trends. Median estimates would bring labor productivity growth close the 3 percent observed between 1996 to 2005, during the deployment of the internet. However, other studies suggest this may be too optimistic. AI-related productivity gains are likely to vary across countries. Korea is well placed to be among the main beneficiaries, thanks to its high level of digital development and strong position in the global semiconductor market.

AI’s productivity effects will partly hinge on the degree and speed of AI adoption, which in turn will depend on costs, as well as the ability to mobilize complementary factors and rethink production processes and business models. AI models are expensive to develop, but applications are also often costly to run. AI’s high energy consumption may threaten sustainability objectives. Deployment may also be constrained by bottlenecks in complementary factors, such as data and skills. OECD research has documented a widening productivity gap between frontier firms and other companies over the past two decades, which AI may widen further if big tech companies can leverage access to data, skills, and computing power to further entrench their digital markets dominance. Thus, monitoring AI competition will be essential to foster innovation and productivity growth.

While many economists expect the adoption of AI to bring a gradual shift in productivity levels over the coming decade, the most optimistic anticipate a permanent increase in productivity growth rates, resulting from accelerated progress in research. Such optimism is partly grounded in the success of some AI applications addressing bottlenecks in research processes. For example, an AI system was able to predict a protein’s 3D structure from its amino acid sequence, opening great prospects for biology and medicine. However, historically, technological progress has tended to come in waves, despite the emergence of ever more powerful research tools. Furthermore, the translation of scientific discoveries into economic applications is never straightforward and is often contingent on factors beyond technology.

Altogether, both fears of large-scale job destruction and hopes of a huge boost to productivity growth seem overblown at this stage of AI development. Nevertheless, AI will transform the way we work, affect business models, and have distributional consequences. Recent studies suggest that AI adoption may reduce wage inequality within professions, as AI seems to raise the productivity of inexperienced workers disproportionately. However, the key issue is whether AI will primarily be used for automation, resulting in job cuts and potentially lower wages for displaced workers, or whether it will bring in new goods and services, new business models and new jobs. In the former case, AI is likely to raise the share of capital income, prolonging a four-decade rise in income inequality across most advanced economies. Even in the latter case, skilled professionals working with AI may benefit more than workers in occupations less prone to AI enhancement, which could also worsen income inequality. Policies need to promote an inclusive deployment of AI, by investing in education, promoting reskilling and upskilling in cooperation with the social partners, facilitating job transitions through job search support and career counselling, ensuring fair competition, limiting differences in the tax treatment of labor and capital income, and protecting consumers and citizens from harmful uses of AI.

Christophe Andre

Christophe Andre is a senior economist at OECD. The views expressed in this column are those of the author and not of the OECD or of its member countries. -- Ed.