March of the machine makes idle minds?
A synthesis of what economics and history suggest about AI and jobs
Each wave of technological change has arrived with a familiar fear: that machines will make workers redundant. In the 1920s, as industrial machinery spread through American factories, newspapers blamed "labor-saving devices" for the prevalence of unemployment. Similar claims have resurfaced repeatedly since, for example, with computers in the 1980s and robotics in the 2000s.
The historical record can temper these fears. Over the past century, despite the introduction of technological innovations, workers in advanced economies have not drifted into mass idleness. In the United States, unemployment has typically hovered around single digits, aside from recessions and the Great Depression. Labor force participation has remained high, with roughly two-thirds of the working-age population either employed or looking for work.
Yet the arrival of artificial intelligence has revived old anxieties about the effects of technological advancements on employment. Today's headlines are announcing corporate layoffs and hiring freezes "driven by AI", with particular concerns for entry-level workers and white-collar jobs. Students are told to rethink their career choices. The recurring question of technology's effect on employment is therefore in full swing. Is this time different?
To answer it, economists typically look to a mix of theory, history and early evidence.
Tasks, not jobs
Economists assess the effects of technology on work at the level of tasks. This task transformation literature considers a "task" as a unit of activity; an occupation a bundle of tasks; and employment or jobs as the number of people employed in that occupation. Tasks are in turn split into routine and non-routine tasks, and cognitive and manual tasks. Routine tasks follow clear rules, such as repetitive assembly work or data processing. Non-routine tasks are more challenging to codify as these rely on judgement, tacit knowledge or interpersonal interaction — think of creating a treatment plan for a patient as a non-routine cognitive task.
Machines affect tasks, occupations and employment through three channels. Displacement occurs when tasks are automated and substituted by machines. Augmentation raises productivity by complementing workers in performing their tasks. Reinstatement reflects the creation of new tasks and occupations altogether. In reality, a combination of these three may occur together. Consider the occupation of the "paralegal", which performs tasks such as document review, contract summarization and legal research. Software that automates the review of documents does not eliminate the paralegal profession altogether; instead, the software reshapes what tasks paralegals perform.
Historically, automation has fallen hardest on routine tasks. Today, AI extends the capabilities of machines into non-routine domains, particularly cognitive tasks such as writing, coding and analysis. This fuels today's concerns that AI's impact on employment may be broader than that of previous technologies.
From predicting technology exposure to assessing actual job impacts
Recent research has leveraged these models to attempt to quantify how exposed tasks, occupations, jobs and wages are to computerization and AI. Estimates vary, with early studies suggesting that nearly half of American employment is at high risk from computerization. More recent work highlights that few occupations are fully automatable, even if many tasks and occupations may be reshaped by computerization, rather than outright eliminating jobs, especially by generative AI since 2022.
Exposure, however, is not outcome. Predictions based on technical feasibility say little about how firms and workers actually embed AI into production. Evidence from early usage confirms that AI is deployed to both automate and augment specific non-routine cognitive tasks such as coding and writing. That fits with the technology's current limitations. It also reflects how firms respond to productivity gains.
When tasks become cheaper due to advances in machine capabilities, firms have options. They may reduce headcount while maintaining output. They may slow hiring while growing output. Or they may expand: Uber's chief executive recently argued that AI makes software engineers more productive, and thus could induce the company to hire more engineers. Such effects complicate simple automation stories.
So far, while recognizing research gaps, the labor-market data show little sign of large-scale disruption. Early findings suggest entry-level workers may face greater labor displacement, as AI substitutes for routine work typically conducted by juniors, while senior workers may benefit more from augmentation. If sustained, this could weaken traditional career ladders as we know them today. But overall employment effects remain difficult to detect; we do not know (yet) to which extent AI exposure translates into labor market impact.
Why the effects may appear muted
One explanation is that AI capabilities, though impressive, remain in the early stages of development. Today's systems excel at narrow tasks ("artificial narrow intelligence"); there are still challenges in the ability of AI to perform tasks reliably, especially for judgement-based and multi-step tasks. Predictions of more sweeping impacts on the labor market often rest on assumptions about future machine capabilities rather than current ones.
Another reason is the speed of diffusion. New technologies rarely transform economies overnight. Firms must first adopt the technologies and then redesign workflows before any impact materializes. Reliable data on firm-level AI adoption remain limited, but evidence suggests that uptake is concentrated among large enterprises and professional services industries.
History offers further caution. As Elting Morison noted in Men, Machines, and Modern Times, technologies rarely fail because they do not work; they fail because institutions struggle to adapt to them. Overcoming these requires complementary innovations, such as changes in skills and managerial practices. Even when AI performs well in controlled settings, translating that potential into impact requires organizational adjustment by firms, which may be slow and uneven.
Cost is another constraint. In areas such as computer vision, recent studies suggest that not all exposed tasks are economical to automate given the costs of GPUs and deployment. In regulated professions such as radiology and accountancy, professional standards can also temper adoption. A further complication is that tasks may be intertwined, which makes them inefficient to be conducted in isolation. Splitting them between machines and humans can introduce coordination costs and affect quality. In one study, software developers took roughly 19% longer to perform coding tasks with AI assistance than without, considering the amount of time spent prompting and rewriting AI-generated code rather than simply writing it themselves.
Finally, exposure estimates focus on task-level feasibility rather than broader general-equilibrium effects. When technology makes the performance of some tasks cheaper or more reliable, it can raise the value of the tasks that continue to be performed by humans, and can thereby adjust the composition of occupations and ultimately increase the demand for labor. For example, although routine cash-handling tasks declined following the introduction of ATMs, the employment of bank tellers rose modestly as banks expanded branch networks and tellers shifted toward customer relationship management. Such economy-wide adjustments take time.
The known unknown: labor reinstatement
Predictions of exposure also often overlook the creation of new occupations and jobs. Many occupations that are now commonplace did not exist one or two generations ago. A recent study suggests that around 60% of occupations present in 2018 did not exist in 1940, which is largely due to demand for new skills induced by technological innovations. The occupation "software developer" is an obvious example.
AI is already starting to generate new forms of work. Tasks such as prompt engineering appear increasingly in vacancies. Occupations focused on data curation, human-AI collaboration and responsible deployment have emerged alongside the recent surge in AI development. These jobs are still small in number and poorly measured, but so were many now-common professions in their infancy.
Striking the right balance
Why, then, does job destruction dominate the conversation? Partly because humans are drawn to worst-case scenarios. Partly because job losses are easier to anticipate and count. We can count jobs that disappear; new occupations take time to emerge and even longer to be recognized in statistics.
None of this guarantees a painless transition. The risks and potential for short-term disruption are present, especially for workers whose roles are most exposed. But history suggests that new machines change what people do more often than whether they work at all. Artificial Intelligence, at least considering its capabilities today, looks set to follow that pattern.
Until then, the evidence points less to idle minds than to reoriented ones.