When New Technology Arrives, Who Wins and Who Loses?

While generative AI like ChatGPT represents a major advancement, it is certainly not the first time a new tool has sat poised to change (or even eliminate) certain jobs. And that previous experience may provide important lessons about what we can expect in the future.

In a new paper, Kellogg’s Dimitris Papanikolaou and Bryan Seegmiller (a professor and assistant professor of finance, respectively) take up the question of how the arrival of new technology affects workers and their earnings. The study was coauthored by Leonid Kogan and Lawrence Schmidt of the MIT Sloan School of Management.

Their analysis focuses on the period from 1981 to 2016—a time span in which many new technologies were introduced across a variety of occupations, both blue- and white-collar.

Perhaps not surprisingly, the researchers found that when a new tool can perform a task in place of a worker, all affected workers suffer. “They experience a loss of wage earnings, and that is largely independent of age, their income level, which sector they’re working in, the type of job that they do, or whether they have a college degree,” Papanikolaou says. But when a new technology complements workers performing a task, the effects are more variable: the most experienced and highly paid workers suffer, while new hires appear to benefit.

Papanikolaou, Seegmiller, and their colleagues also studied the potential ramifications of AI on today’s workers. They found that “AI, as a technology, levels the playing field within an occupation,” Papanikolaou says. In other words, if everyone can code, a skilled and experienced coder will be less valuable in the job market. The upshot is that “it’s going to hurt workers that are better at their jobs.”

Measuring exposure to new technology

The researchers began by gathering descriptions of jobs from ONET, the Occupational Information Network, as well as the 1991 edition of the Dictionary of Occupational Titles—both widely used sources of information about different professions and the functions they perform. For example, the ONET entry for the job of “Kindergarten Teacher, Except Special Education” lists 37 tasks, among them “Demonstrate activities to children” and “Read books to entire classes or to small groups.”

Then—in a perhaps ironic twist—the researchers asked ChatGPT to classify each job’s tasks as either routine (requiring little experience and likely easy to automate) or nonroutine (requiring lots of experience and probably difficult to automate). They also validated ChatGPT’s results against other classification methods to ensure its accuracy.

For example, as a professor, “sometimes I teach, sometimes I write papers, sometimes I field reimbursements,” Papanikolaou explains. Dealing with reimbursements “is probably routine, while the other two are totally nonroutine.”

Next, they gathered a list of important patents issued from 1980 to 2007. They focused on so-called breakthrough patents—those that were very different from previous patents but very influential on future patents.

Then, the team calculated how closely these breakthrough patents matched routine and nonroutine job tasks. If a patented technology was closely related to a routine task, the researchers considered it a labor-saving technology—one likely to fully automate that task. If a patented technology was closely related to a nonroutine task, the researchers considered it a labor-augmenting technology—one likely to complement a worker performing that task.

This information was used to compute measures of exposure to labor-saving and labor-augmenting technology for different occupations. A job experienced high exposure to labor-saving technology if many of its tasks were automated within a given time period; similarly, a job experienced high exposure to labor-augmenting technology if many of its nonroutine tasks were complemented by technology within a given time period.

How technology exposure affects occupations and workers

To determine the consequences of exposure to labor-saving and labor-augmenting technologies, the researchers gathered U.S. government data on workers in different occupations, including their earnings, ages, and levels of education from the years after the breakthrough patents were granted.

Overall, for any given occupation, exposure to labor-saving technology predicted lower wages and lower employment. Exposure to labor-augmenting technology, meanwhile, predicted higher wages and higher employment for that job.

But when the researchers shifted gears and began analyzing the effects of technology exposure at the worker level, rather than the occupation
level, they discovered a more complicated story—particularly when it came to labor-augmenting technology.

Across all occupations, the average worker whose exposure to labor-augmenting technology suddenly spiked saw a small decrease in earnings and a small increase in the likelihood of losing their job. These trends were even more pronounced among white-collar workers, older workers, and highly paid workers within an occupation.

This finding, combined with the knowledge that wages increased at the occupation level, “leads you to suggest that a lot of the benefits go to newly hired workers,” Papanikolaou says. In other words, workers who were used to doing things a certain way struggled to adapt when complementary technology arrived, while less-experienced workers could harness the power of these new tools.

What does the AI future hold?

But will the trends of technological advancement in decades past extend to AI? To find out, Papanikolaou, Seegmiller, and their colleagues ran a new analysis. This time, instead of using patents as their proxy for technological change, the researchers asked ChatGPT whether AI could perform a given task without human intervention or whether the task would require significant human intervention.

As before, the researchers matched this information with occupation descriptions to determine whether a job’s routine and nonroutine tasks could be complemented or substituted by AI. They used census data on Americans’ earnings to predict the effect of AI exposure on wages.

While speculative, the findings suggested that workers in office and administrative occupations, as well as those in production and transportation, face high levels of exposure to AI as a labor-saving technology.

Meanwhile, the occupations most exposed to AI as a complementary technology included “Insurance Underwriter,” “Medical Transcriptionist,” “Customer Service Representative,” “Personal Financial Advisor,” and “Budget Analyst.” While less-experienced workers in these roles may benefit from the robotic assistance, the researchers suggest, older ones may struggle and see their wages decrease, because their expertise and experience may not provide the competitive edge it once did.

Complex effects

Papanikolaou says the research offers much-needed perspective on how AI—or any new technology—can change the workforce.

“Especially with the case of AI, everyone freaks out, and they interpret the statement, ‘AI will affect my job’ as ‘AI will do my job for me.’ And two things are not the same, because AI can be a tool or substitute,” he says. “And just because the job may change doesn’t mean that the job will be eliminated.”

The findings also suggest that soft skills may become ever more important in the workforce. In fact, jobs that mostly rely on interpersonal skills were hardly affected by technology at all.

For instance, Papanikolaou points out, AI tools may, in the future, aid doctors and nurses in determining medical diagnoses—but they probably won’t be able to offer the emotional care and support that sick patients need: “That’s something where AI would probably do much worse.”