Earlier this year, the National Academies of Sciences (NAS) released “How Is AI Shaping the Future of Work,” a video podcast and accompanying transcript of a conversation between MIT economist David Autor and writer Sara Freuh in Issues in Science and Technology, a quarterly journal jointly published by the NAS and the University of Arizona.
“For as long as people have speculated about the development of artificial intelligence, they have debated its potential impacts on the labor market,” noted the article in its introduction. “Today, several years into widespread use of large language models, those questions are more urgent, but the answers are less clear. Is AI already taking jobs away? Could human beings flourish in a world in which they no longer have to perform economically valuable work?”
“As artificial intelligence becomes ever more advanced and capable, it’s being harnessed more frequently in American workplaces, and it’s raising thorny questions about how it will impact the nature and number of human jobs in the future,” Freuh added. “I’m joined today by David Autor, a professor at the Massachusetts Institute of Technology and one of the world’s leading labor economists. He’s known for his research on the effects of globalization, automation, and technological change on the labor market … and he’s written extensively on the subject.”
I didn’t focus much on economics during my long career as a technologist and business strategist at IBM. But, since becoming affiliated with MIT in 2005, I’ve become increasingly interested in economics — in particular, on the impact of advanced technologies like the internet and AI on jobs and the labor markets. And, I can personally vouch for Freuh’s comments about Professor Autor’s excellent articles, from which I’ve learned so much over the past couple of decades.
Let me highlight some of the key themes in the conversation between Freuh and Autor.
AI Is Spreading Fast, but Labor Markets Haven’t Collapsed
“There are a lot of news articles, for example, on how many entry-level coding jobs are going away and speculation about how AI is impacting entry-level work in general,” said Freuh. “Do we have good big-picture data on how many workers are using AI in their jobs and the degree to which workers are being replaced by AI so far?”
Autor replied that AI has caught on incredibly quickly. At least half of workers are using AI in their jobs, many even on their own, without their employer’s knowledge. “It’s used at home, it’s used at work, it’s used by people of all ages, and it’s used now equally by men and women and across education groups. So it’s pretty broadly used.”
However, we still don’t know how many jobs have been lost to AI. The general labor market has been slowing since April 2022 — before the release of ChatGPT in November of that year. Recent articles have identified AI-exposed occupations like software development and customer service as experiencing declining employment among recent college graduates. But, employment growth among young people has declined in every sector regardless of degree, so it’s not clear whether the slowdown can be attributed to AI or to broader economic factors.
He also reminded us that most research on the impact of technology on jobs has focused on automation, while being relatively silent on the countervailing emergence of new occupations that augment human labor. For example, data scientist is a relatively new occupation that wasn’t recognized by the US Bureau of Labor Statistics (BLS) until 2018, and there are now over a quarter million data scientists in the United States.
The Devaluation of Expertise
Freuh noted that part of the anxiety in the workforce stems from the fact that we can envision jobs that may disappear due to AI, but we’re still unclear about what kinds of new AI-based jobs will be created.
“Let’s say a million jobs are destroyed, and a million are created,” said Autor. “The people who are taking the new work are not usually the people displaced from the old work.” Even if the labor market ends up “5% better on average,” he noted, it might be “90% worse for some people and 95% better for others.”
“People are paid not for their education, not for just showing up, but because they have expertise in something. It could be coding an app, baking a loaf of bread, diagnosing a patient, or replacing a rusty water heater. When technology automates something that you were doing, in general, the expertise that you had invested in suddenly doesn’t have much market value.”
For example, taxi drivers once had highly specialized knowledge of streets and routes that was worth a great deal. That expertise has been devalued by smartphone navigation apps now available to anyone. Similarly, language translation, once a high-level cognitive skill, can now be done quite well by AI-based applications.
In a February 2024 article, “AI Could Actually Help Rebuild the Middle Class,” Autor defined expertise as the domain-specific knowledge or competency required to accomplish a particular goal. Expertise commands high wages if the goal it enables is both necessary and relatively scarce — for example, physicians, engineers, and lawyers. In contrast, jobs that require little expertise and training generally command lower wages — such as waiters, janitors, and school crossing guards.
AI, he argues, offers the opportunity to extend the value of human expertise by enabling “a larger set of workers equipped with the necessary foundational training to perform higher-stakes decision-making tasks currently arrogated to elite experts, such as doctors, lawyers, software engineers, and college professors.”
Rapid Transitions in the Labor Market
Autor added that economic transitions are always costly unless they happen slowly, because occupational change is usually generational. We saw this in the early 2000s, when more than a million manufacturing jobs were lost as production moved to lower-cost countries, especially China.
In a May 2011 article, “The Polarization of Job Opportunities in the US Labor Market,” Autor wrote about the changing dynamics that had led to a sharp rise in inequality. The structure of job opportunities in the United States had polarized, with expanding employment in both high-skill, high-wage occupations and low-skill, low-wage service occupations, coupled with contracting opportunities in middle-wage, middle-skill white-collar and blue-collar jobs. The decline in middle-skill jobs proved particularly detrimental to the earnings and labor force participation of workers without a four-year college education.
“More than a million manufacturing jobs were lost,” he told Freuh. While that number was not large relative to the scale of the US economy, the losses were highly regionally concentrated in the South Atlantic and the Deep South. A relatively small number of counties saw their lifeblood industries wiped out. Many displaced workers had to change locations or careers. The manufacturing work they had performed required specialized expertise and experience; the next available jobs were often lower-paid service positions — food service, cleaning, janitorial work, home health aide, security — that did not command comparable wages because they required less specialized expertise.
Policy Responses
“Do you think that we are potentially facing something of that magnitude again, but this time mostly with knowledge workers rather than manufacturing jobs?” asked Freuh.
“The greatest similarity is that this could happen quickly in certain areas, in certain activities, and it’ll be extremely disruptive and scarring for the people who lose that work and have to do something that’s lower paid and not as consistent with their skillset,” replied Autor.
But there are important differences. The loss of blue-collar jobs in the 2000s was regionally concentrated — a relatively small number of manufacturing communities lost a large number of jobs. By contrast, white-collar clerical and administrative jobs are spread throughout the country, so AI is unlikely to “knock out an entire community all at once,” thus the pain may be more broadly distributed.
One similarity, however, is the disruptive and potentially long-lasting consequences of job loss. This time, Autor argued we should have stronger safety net systems in place. Wage insurance, for example, is a promising idea for workers displaced by AI.
Most important, “we should be thinking about shaping where we want to direct the technology, not just adapting to what happens.” AI could be used to expand expertise in areas like healthcare, where there are worker shortages; to help teachers become more effective by providing better tools; and to support lifelong training and skill development.
In my opinion, the key takeaway from Autor’s analysis is that AI will reshape — and potentially devalue — human expertise in uneven and disruptive ways. The long-term question is less whether AI will destroy jobs and more how it will redistribute the value of skills, occupations, and opportunities across the labor market.
Cautious, Responsible Optimism
Freuh finished their conversation by asking Autor whether he was hopeful that we will learn to navigate AI in a way that helps most of humanity.
“I think one should distinguish between what I think is possible and what I think is likely,” replied Autor. What’s possible is substantial. There is enormous opportunity. Over the past four decades, global welfare has improved dramatically — not only in China, but across Sub-Saharan Africa and Latin America. Technological and economic development have delivered substantial gains in prosperity.
Yet optimism must be paired with realism. The right attitude, Professor Autor suggested, is to be both optimistic and pessimistic simultaneously — to pursue opportunity while building systems that help society manage risk.
We’re not passive observers of technological change. AI will shape the labor market, but the direction it takes — and who benefits — will depend on policy choices, institutional design, and deliberate effort.
If managed well, AI could broaden access to valuable expertise and increase prosperity. But there is nothing automatic about equitable outcomes. The ultimate test will not be whether AI advances, but whether we shape its deployment in ways that expand opportunities rather than deepen divides.