By some estimates, one out of every 13 working women in the United States in 1950 was employed as a telephone switchboard operator. But with the rise of automatic switching systems in the 1960s, the job all but disappeared.
A decade later, as computers gradually moved to the center of American working life, the ranks of information technology workers, such as computer programmers, began to swell—from 450,000 in 1970 to 4.6 million in 2014.
Will artificial intelligence make us all telephone switchboard operators or IT professionals? Around the world, doomsayers and optimists are duking it out over this very question.
The debate isn’t surprising to Menaka Hampole, an assistant professor of finance at Yale SOM. She says that understanding what any new technology means for the future of a particular occupation is fiendishly difficult.
“Firms do different things for different reasons, and it can affect different occupations differently,” she explains. These idiosyncratic decisions by companies that are dissimilar to begin with can make it hard to know precisely what role the technology is playing in workers’ fates.
A new working paper by Hampole—alongside Dimitris Papanikolaou and Bryan Seegmiller of Northwestern University and Lawrence D.W. Schmidt of MIT—attempts to disentangle this knot of differences. Hampole and her colleagues used real-world data and a theoretical model to pinpoint the effects of AI on demand for specific occupations and jobs as a whole.
The results were nuanced. Demand for some occupations that were heavily exposed to AI did decrease. At the same time, other occupations exposed to AI actually grew, seemingly because AI made those workers more effective. And overall, the researchers found, firms using AI ultimately became more productive and were able to expand their workforces. In other words, AI shifted the kinds of jobs on offer, while task reallocation and growth at AI-adopting firms offset much of the decline in demand for more-exposed occupations.
To Hampole, this finding complicates popular preconceptions about AI. It’s not necessarily the case that firms adopt AI and simply fire workers. AI can replace labor on some tasks but workers can shift effort toward other tasks, and AI adoption can also make firms more productive and lead them to expand employment.
The researchers started by analyzing—with the help of AI, naturally—millions of LinkedIn profiles from 2014 to 2023. Most of this period predates the widespread adoption of generative AI; the AI applications in the data primarily reflect an earlier wave of machine-learning tools used for prediction, classification, recommendation, recognition, and related tasks. For the first time, tasks such as fraud detection, product recommendation, and speech recognition could be wholly or partially outsourced to computers.
Looking through the LinkedIn profiles allowed the researchers to glean information about which firms were using AI and in what ways. For example, they found that firms that adopted AI were larger, better-paying, and more productive than those that didn’t—likely because it takes money and expertise to implement such systems to begin with.
Then they compared how closely the functions performed by AI matched job descriptions in O*NET, a database that compiles information about the tasks associated with different occupations.
For example, one JPMorgan Chase employee wrote in their LinkedIn profile that their responsibilities included “AI/ML [machine learning] model delivery” and the “development and deployment of quantitative risk models that serve regulatory and credit risk assessments.” Textual analysis revealed that this use of AI closely resembles one of O*NET’s listed responsibilities for credit analysts: “Analyze credit data and financial statements to determine the degree of risk involved in extending credit or lending money.”
The researchers call this relationship “exposure.” When a job task matches something an AI tool can do, that task is considered highly exposed to AI, and is likely to be replaced by it. Indeed, the researchers found that tasks that were highly exposed to an AI tool developed by a company in a given year were significantly less likely to be mentioned in future job postings by the same firm. For example, job descriptions at JPMorgan Chase were less likely to mention credit risk analysis among their duties following the deployment of an AI tool that could do the same thing.
Interestingly, the researchers discovered, many of the most AI-exposed tasks were performed by highly paid occupations, including market research analysts, computer hardware engineers, and logisticians.
But just because a task can be done by AI doesn’t necessarily mean an entire job can be. “AI doesn’t hit jobs unilaterally,” Hampole explains. If a worker is responsible for 10 tasks, “the question is, can AI do all of those tasks, or can it only do 2 of her tasks?” When AI exposure is highly concentrated across just a few of a worker’s tasks, rather than spread more evenly across all of their tasks, that job is less likely to be displaced.
Why? Because with, say, two tasks taken entirely off your plate, “you have more time to reallocate your efforts toward the things that AI can’t do,” Hampole says—potentially making you even more productive. This concentration phenomenon is “a really important force,” Hampole says, and one that is easy to overlook.
What’s more, AI adoption seemed to help firms hire more workers overall. After controlling for a variety of potentially confounding factors, the researchers discovered that the more firms integrated AI, the higher their revenues, profits, productivity—and, crucially, employment.
These two phenomena—exposure concentration and the effects of AI on firm growth—help explain the complex patterns that emerged when the researchers modeled the effects of AI on demand for different occupations. They looked specifically at how AI influenced the employment share (that is, the portion of all jobs held at the time) for each occupation in their data set.
There were notable winners and losers: for example, occupations in business, financial, and engineering fields saw declines of almost 2% in their employment shares over five years. And AI was an important overall driver of changes in the labor market, explaining about 14% of the variation in occupational employment-share growth over the sample period. But once task reallocation and AI-driven firm growth were taken into account, the estimated effects of AI on the composition of employment across occupations were modest. The net effect reflects two offsetting forces: workers can reallocate effort towards less-exposed tasks, and AI-adopting firms can grow and increase labor demand.
Although the research was conducted before generative AI tools like ChatGPT became available, Hampole says it’s an open question whether the same patterns will hold in the generative-AI era. However, the same underlying forces—direct substitution, workers reallocating across tasks, and productivity-driven firm growth—are likely to remain relevant.
For now, she hopes the current study sheds light on the nuanced and often unexpected ways that new technology can influence the labor market. “Popular media focuses on the fact that firms can use AI to cut labor,” she says—and that’s true, but it’s not the only dynamic at work. “Firms can become more productive as a result of AI, and in the economy at large, that can actually increase employment. All of these different forces matter.”
“The Yale School of Management is the graduate business school of Yale University, a private research university in New Haven, Connecticut.”
Please visit the firm link to site