Minimizing the negative effects of technological unemployment in the age of AI:A realist view
By Janine Berg, Senior Economist, International Labour Organization
Current debates on Artificial Intelligence (AI) and jobs have centred on two opposing viewpoints: the pessimists, who fear widespread unemployment and a future without work, and the optimists, who see new technology as the means to relieve workers of mind-numbing tasks, and where vast productivity gains will usher in a richer and more glorious future.
But there is also space for an intermediate position, which acknowledges the risks as well as the potential rewards. Let me call them the realists.
The realist view, first and foremost, acknowledges that the outcomes are not set in stone. Societies can decide how, and if, technology is deployed, how the possible gains are distributed, and what happens to those affected, for better or for worse.
It recognizes that most jobs won’t disappear as there are limits to what AI can do, and even greater limits to what it can do well. But it also recognizes that there will be some job losses and that the consequences for the workers who do lose their jobs are not pretty, both in terms of the immediate unemployment effects but also future employment and earnings.
“Operators who were made redundant were more likely to be unemployed when compared with their peers, and if they did find a new job, it was likely to be lower paying. ”
Economic history is full of stories of hardship suffered as a result of technological innovation. Writing on technological unemployment in the Industrial Revolution, the historian Ben Schneider documents the negative long-term effects for both women and their families caused by the mechanization of hand spinning. In the 1770s, hand spinning in Britain provided work for more than eight percent of the population, primarily women and – in those days – children. The loss of this home-based work, commencing in the 1780s and persisting for half a century, reduced rural incomes, as the women were not able to substitute the income loss. The new factory jobs that did emerge were in urban centers and were far fewer: in 1850, such employment accounted for less than one per cent of the population, with fewer than half of the jobs occupied by women and girls.
The mechanization of telephone switchboards is another example. In the 1920s, the US telephone industry employed over 300,000 people, and was the fifth most important occupation for young women. Mechanization, much of which occurred during the 1920s and 1930s, led to an 80 percent drop in employment. While the elimination of these entry-level positions did not negatively affect women entering the labour market, operators who were made redundant were more likely to be unemployed when compared with their peers, and if they did find a new job, it was likely to be lower paying.
“we shouldn’t ignore the negative consequences of technological unemployment in the short to medium-term”
While we know that eventually these technological innovations, but also other inventions in shipping, transport, digitalization and other areas, were beneficial for economic growth and employment expansion overall, we shouldn’t ignore the negative consequences of technological unemployment in the short to medium-term.
Our research at the ILO suggests relatively small employment losses from generative AI, but effects that will nonetheless be concentrated, particularly among clerical support workers (see Figure 1). Clerical support workers includes professions, such as customer service workers, receptionists or secretaries, that have experienced declining employment levels over the past 10-15 years, and where the effects of AI are just beginning to take hold. Many of these clerical support jobs are held by women. As a result, women are 2.5 times as exposed to automation risks than men. Overall, we estimate that 2.3 per cent of employment (or 75 million jobs) is at risk of automation because of high exposure to generative AI technology. In high-income countries, the share is greater at 5.1 percent of employment (or 30 million jobs), as this type of work is more prevalent.
Figure 1. Tasks with medium and high exposure to Generative AI, by occupational category
Note: Occupational categories at ISCO-08 1-digit level. Levels of exposure to potential automation by GenAI with capabilities similar to GPT-4 on 0-1 scale. “Medium exposure” for 0.5-0.75 scores and “high exposure” for scores greater than 0.75.
Source: Gmyrek et al. 2023.
Also troubling is that there is not much of a buffer to the risks of automation, even in some developing countries. A recently published ILO-World Bank study by my colleague, Pawel Gmyrek and his co-authors, finds that in Latin America, many of the occupations that could benefit from the productivity-enhancing effects of AI do not currently use a computer at work, and thus will miss out on these benefits, whereas workers in jobs at high risk of automation are, for the most part, using computers (see figure 2). Thus, inadequate infrastructure is a bottleneck to productivity gains in some occupations, but not in those at risk of automation.
Figure 2. Exposure to Automation and use of computer at work, Latin America
Source: Gmyrek et al. (2024).
In Latin America, the at-risk jobs are more likely to be held by women who are relatively well educated, living in urban areas, with a relatively high income and a formal, salaried employment contract in the banking, finance and insurance industries, or in the public sector. In other words, they are pretty good jobs. And while not all jobs will be shed, those who do lose their jobs will struggle recovering, especially in Latin America where the labour market is comprised of high shares of informal work.
The literature is clear that unemployment, whether for technological or other reasons, inflicts a longer term “scar” on workers, both in the likelihood of recurring bouts of unemployment, but also in lower subsequent earnings over time. The findings hold regardless of the country, the economic cycle, or the characteristics of the worker. Study after study, each more sophisticated than the last, confirms the result.
But the studies also show the importance of transfer payments in reducing immediate income loss as well as in reducing scarring effects by allowing workers the time to search for re-employment that is of good quality. For this reason, scarring effects are less severe, though still evident, in countries with more robust labour market institutions and social protection systems.
Which is why policies are so important.
“...encourage technological innovation that complements human labour rather than replaces it...”
The first, best solution is to avoid job loss. One way to do this is to encourage technological innovation that complements human labour rather than replaces it, as advocated by Acemoglu and Johnson in their recent book, Power and Progress. Another option – that is more feasible in the short-run – is to redeploy those staff at risk of technological unemployment to other jobs within the same organization. In addition to redeployment, the ILO’s Termination of Employment Recommendation, 1982 (No. 166) encourages employers to explore other solutions, including through adjustments to working hours and hiring policies, and to undertake such measures after consultation with employers’ and workers’ organizations.
In the event of job loss, other policies are needed, including income support through unemployment insurance or other social protection measures. There is also a need to develop and institute reskilling and upskilling programmes that can prepare workers for new careers in the digital economy, as well as in the growing green and care economy. Preparing workers for these jobs and supporting public and private investments in these sectors will go a long way to minimizing the negative effects of technological unemployment.
This article has been pre-published with permission, on the occasion of the launch of the ILO Observatory on AI and Work in the Digital Economy.