Will gen AI cause job loss? It depends on whether it augments or automates human labor, a new MIT study suggests.

July 23, 2024

New technologies can profoundly affect work. They can alter occupations and create new ones. They can generate or eliminate demand for specific skills. They can lift or lower employment. While most of these effects have been well documented, economists have struggled to trace how new technologies create work—until recently.

A new study by MIT economists associated with the Shaping the Future of Work Initiative provides compelling evidence that innovations that augment—or enhance—human labor tend to create new work, but innovations that automate tend to slow the creation of new work.

The groundbreaking research provides new theoretical insights into the relation between technology and work. But more practically, the research offers evidence-based ways to think about emerging technologies like generative AI, an innovation that is having uncertain effects on the quantity and quality of work available to humans.

The research, published in The Quarterly Journal of Economics, helps solve a longstanding economic puzzle. According to the study’s authors, economists had long known that automation tends to take away jobs—a perhaps unsurprising result. But economists had lacked data showing the opposite, that technologies create jobs, despite the fact that they are widely believed to do so.

The research developed a new method “to assess whether … new work [created by innovations] exerts a countervailing force to the employment-eroding effects of task-displacing automation,” write the researchers led by David Autor, MIT professor of economics.

The researchers analyzed patent and census records and data from 1920 to 2018, looking for evidence that new work (opposed to more of existing work) emerged and from where that work originated. The study used natural language processing tools to trace new work to patents issued throughout the period. In other words, the researchers found a way to trace how new inventions, represented by patents, resulted in new occupations and occupational tasks, evidenced by Census records.

The research finds that, during the period studied, new work emerged in the U.S. from two sources: augmentative innovations ( “technological innovations that complement the outputs of occupations”) and demand shocks that raised occupational demand. In contrast, automation did not yield new work but decreased labor demand and even slowed the emergence of new work.

In addition to finding that augmentative tech increases new work, the researchers were able to document distinct shifts in the kinds of new work created during the twentieth and early twenty-first centuries. The results suggest that automation innovations have had the greatest influence in recent years and are at least in part responsible for current wage structures. Wages and occupations have been polarizing since 1980.

The study found that “the focus of new work creation has shifted from middle-paid production and clerical occupations to high-paid professional and, secondarily, low-paid services since 1980,” write the researchers. Current evidence suggests “that the last four decades have seen relatively more automation and less augmentation than the prior four.”

These results hint at many enticing future directions, for both policy and research.

Because augmentation innovations tend to lift both employment and wages, the researchers suggest that “new work” stemming from augmentations may be more valuable than simply “more work.” If “new work” adds more value than “more work,” as the authors propose, “then policies that foster new work creation may be of particular interest.”

But the most pressing question may be whether generative AI, like Chat GPT, should be classified as an augmentative technology—thus creating new work—or an automating technology—thus slowing new work and potentially exacerbating current occupational polarization. 

“Early evidence on this question points to a broad range of potential effects” for generative AI, write the researchers. “But definitive findings are elusive.”

Back to All News & Research