Why AI hasn’t caused a job apocalypse — so far

Over the past few months, several surveys and media reports have highlighted how artificial-intelligence technologies will increasingly displace workers. And a growing number of companies mention AI as a factor in planned or actual lay-offs. For instance, data from Challenger, a recruitment firm in Chicago, Illinois, that tracks companies’ public announcements, suggests that, in 2025, AI might have been responsible for seven times as many lay-offs in the United States as were the international tariffs the US government have imposed, which are currently a major source of economic disruption globally.

Such warnings have fuelled widespread expectations that the labour market might be on the verge of upheaval. The most drastic forecasts compare the current moment with the First Industrial Revolution (1760–1830), when mechanization ultimately improved living standards but caused widespread disruption to workers’ livelihoods over the short term.

AI is threatening science jobs. Which ones are most at risk?

Knowledge workers and even researchers like me are wondering whether our jobs might be on the line. Will AI take over teaching, research tasks and data collection? Will the workforce need fewer humans in the future?

These fears are heightened because of speculation that the AI revolution could unfold much more quickly than did previous technology-fuelled transformations. The invention of electricity, for instance, did little to disrupt labour markets on its own. The upheaval came later, through the applications it enabled: indoor lighting, elevators, microphones and other tools reshaped work. Although there were winners and losers, large segments of the workforce had years — sometimes decades — to adapt. Will AI challenge this precedent?

Available data suggest that we shouldn’t panic about the downsides of rapid AI adoption just yet. Research by my team at the Budget Lab at Yale University in New Haven, Connecticut, and the Brookings Institution in Washington DC has found no evidence so far that employment patterns have begun to shift meaningfully since the launch of the ChatGPT chatbot in 2022 (see go.nature.com/47aeexk). The upside is that academics and policymakers still have time to work this problem out — and to identify who might need support during the transition. At the same time, the findings underscore the urgent need for better data-collection systems to study the current transformation.

Poor adoption

Despite a flurry of public announcements by chief executives about imminent AI-driven transformations, not all companies are using machine-learning tools in their day-to-day operations. Survey data collected by the US Census from mid-2023 to February 2026 to track short-term technology use show that about 18% of businesses currently report using AI in the preceding two weeks, and just 22% expected to do so in the next six months (see go.nature.com/3prvpjg).

These figures might overstate effective adoption. Executives are often under pressure from shareholders to articulate an ‘AI strategy’ and to seem forward-leaning. However, AI models might be irrelevant to a firm’s operations, they might lack a compelling application or the necessary tools haven’t matured yet.

AI is transforming the economy — understanding its impact requires both data and imagination

Such a dynamic is not unusual for the early stages of a technological transition. The point is not to dismiss the change, but to avoid being swept up in business leaders’ empty rhetoric. Counting how many times chief executives mention AI during shareholder meetings is, frankly, not very useful.

What might be a more helpful signal? When an innovative technology reshapes the labour market, it eventually leaves a clear fingerprint: employment levels in some occupations contract, but expand in others. For example, since the start of the computer and Internet eras, the demand for computer programmers has risen, whereas the need for some roles, such as secretarial ones, has decreased. For instance, according to the US Bureau of Labor Statistics, about 1.5 million executive assistants were employed in the United States in 2007; by 2023, that number had fallen to below 500,000.

These jobs did not vanish overnight, but the transition had profound consequences for workers with skills relevant to older technologies. It would be a mistake to dismiss the real strain that job transitions impose on workers. One study, for example, estimated that roughly 60% of people’s jobs in 2018 comprised occupations that did not exist in 1940 (D. Autor et al. Q. J. Econ. 139, 1399–1465; 2024).

Therefore, if generative AI tools are already transforming the nature of work, there might be a noticeable drop in the number of administrative assistants employed, as well as a marked rise in AI-related roles, such as data-engineering or machine-learning specialists. So far, however, job distributions are changing at no greater pace than they did before 2022 (see ‘AI adoption footprint’).

Similarly, unemployment periods might be expected to become longer for people with jobs that are vulnerable to replacement by AI, such as translators or legal secretaries. But that’s not yet evident either.

New data system

For now, much of the discussion about how AI might affect the workforce remains speculative. Many of the people most anxious about their futures are young workers, yet there is little definitive evidence to indicate which groups will be most affected. Poor targeting of government-support schemes has real risks: policies designed to protect one segment of the workforce could miss those who ultimately are hit the hardest.