Commentary · AI & Work

Jon Stewart’s AI Jobs Argument Is Missing Part of the Story

The executives deserve the scrutiny. The economics deserves more than a punchline.

By Jeff FergusonOctober 26, 202611 minute read

I caught a Jon Stewart clip on TikTok the other day about artificial intelligence and jobs. I laughed. Then something about it kept bothering me.

Stewart was doing something he does exceptionally well: putting powerful people’s promises next to their inability to explain how those promises would work. AI executives tell us the technology will transform employment, create new opportunities, and make everyone better off. Ask them what those opportunities will be, and suddenly the answers get less impressive.

That is a fair target. People selling a transformation of the economy should expect questions about the people living in that economy.

In the jobs portion of his October 5 AI segment, Stewart told the executives to acknowledge that “if this works the way you say it will work, there aren’t going to be that many jobs anymore.”

The qualification matters. He was challenging their own extravagant descriptions of what AI could do. If somebody promises machines capable of performing virtually everything people do, asking where human employment fits is entirely reasonable.

But the montage also encouraged a broader inference: these executives cannot give convincing examples of future jobs, so their uncertainty exposes the emptiness of the whole argument for new employment. That is where the joke started doing more work than the economics. An awkward answer is not evidence that the economy cannot create work.

I am a giant Stewart fan. I think he is one of the smartest and funniest people doing this. That is why I want to be precise about what bothered me, and fair about what I found when I looked into it.

My first reaction was that he was asking technology executives to answer a labor economics question. That thought was too broad. Stewart interviewed MIT economists David Autor and Daron Acemoglu on April 22, specifically about AI and work. On October 7, he spoke with labor economist Kathryn Anne Edwards.

He has consulted people who study this. My criticism concerns the framing of this particular segment, which offered less of the complexity than his other work demonstrates he can handle.

The longer answers complicate the joke

When Nvidia’s Jensen Huang offered wellness centers and spas as examples of new industries, Stewart had an obvious opening. Spas are hardly a recent invention. It was a bad example, and the joke landed.

But jobs do not have to carry newly invented titles to represent additional employment. A growing economy can support more nurses, electricians, restaurant workers, therapists, or fitness instructors. The occupations can be old while demand for their services grows.

In Huang’s longer interview with Ezra Klein, he makes an argument about human ambition and expanding activity: people use additional capabilities to attempt more things. Klein presses him on whether that adequately addresses workers’ vulnerability. Huang’s confidence remains open to challenge. But the fuller conversation contains an argument about expanding demand that the spa exchange does little to explain.

The same issue appears in Sam Altman’s remarks. In the Economic Times’ edited excerpts of its original conversation with him, Altman discusses adaptation over generations, the speed of the transition, the difference between tasks and jobs, and the possibility that greater productivity will allow people to do more.

Generations is an awfully long answer for someone worried about next month’s mortgage. That deserves scrutiny. Still, difficulty imagining future occupations is not evidence that employment growth cannot happen.

Nor does uncertainty let the executives off the hook. If they volunteer confident claims about net job creation, they owe us an explanation. The distinction is between challenging their confidence and treating their inability to supply a satisfying list as an answer to the underlying economic question.

We have studied what happens next

Autor wrote a paper in 2015 with a title that almost answers the question itself: “Why Are There Still So Many Jobs?”

His explanation includes several forces that get lost when the discussion stops at machines replacing people. Technology can make workers more productive. Lower production costs can reduce prices and expand demand. Higher incomes can support additional spending elsewhere. People also develop new tasks and services around new capabilities.

None of that guarantees a favorable outcome. It explains why counting the tasks a machine can perform does not settle the employment question.

Think about a business that can serve customers with fewer hours of labor. It might reduce staffing. It might lower prices, attract more customers, and expand. It might direct workers toward something it previously could not afford to do. The result depends on demand, competition, management decisions, and the work that remains.

Consider a hypothetical marketing agency. If a tool makes analysis faster, does the agency cut the people performing it, give clients more analysis, serve additional clients, or spend more time deciding what the findings mean? Different businesses can make different choices. An agency with a fixed customer base and little room to grow faces different incentives from one turning away business.

A productivity improvement creates possibilities. It does not choose among them.

Difficulty naming future occupations does not, by itself, establish that employment growth is implausible. We can study the mechanisms that generate employment without knowing every business somebody will build around a technology. We can also recognize those mechanisms without pretending they will operate quickly enough, or strongly enough, to protect everyone.

The history is more useful than either side’s slogan

Economist James Bessen examined two centuries of American textiles, steel, and automobile production in “Automation and Jobs: When Technology Boosts Employment.”

Early productivity improvements helped those industries expand employment because cheaper goods generated substantial additional demand. Producing each unit required less labor, but people bought enough more units to increase total labor demand.

Later, as markets matured and demand became less responsive, further productivity gains reduced industry employment.

That is a more useful lesson than saying automation always creates jobs or always destroys them. The same process can have different effects at different stages. Being able to produce a car more cheaply means something different when millions of households are buying their first one than when most potential customers already have one.

Agriculture provides an even larger example of structural change. According to the USDA’s historical review, agriculture employed 41 percent of the American workforce in 1900 and 1.9 percent in 2000. Mechanization and other productivity improvements helped transform an economy in which farming occupied an enormous share of working people.

That change did not leave a corresponding share of Americans permanently without work. But it occurred over a century, alongside industrial expansion, education, migration, changing households, and new generations entering different occupations. It was not a universal retraining success story.

Even the cotton gin needs care here. I initially thought of it as another obvious example of labor-saving technology. The National Archives’ account makes clear why it cannot serve as a cheerful illustration: automating seed removal made cotton cultivation more profitable and helped expand slavery. It demonstrates that reducing labor in one task can increase demand elsewhere. It also demonstrates how little a rising demand for labor tells us, by itself, about human welfare.

Employment totals are only part of the story.

The economy can adapt while workers get hurt

One of the most revealing studies concerns a job that automation largely eliminated: the telephone operator.

James Feigenbaum and Daniel Gross studied the spread of mechanical telephone switching between 1920 and 1940. Automation eliminated most of the affected operator jobs. Subsequent generations of young women found work in other clerical and service occupations, and their overall employment did not decline.

The women already working as operators had a harder experience. A decade later, they were more likely to be in lower-paying occupations or no longer working.

Both findings matter. The economy developed other employment. The people displaced did not necessarily recover.

Research on industrial robots presents another useful contrast. Acemoglu and Pascual Restrepo found negative employment and wage effects in the United States. A study of Germany by Wolfgang Dauth and colleagues found manufacturing job losses offset by additional service employment, with many incumbent workers taking on different tasks inside their existing companies.

These studies concern different economies and periods. They do not tell us that one policy explains everything. They tell us the outcome is not determined by the machine alone.

For someone losing a career, the eventual creation of employment somewhere else is an incomplete answer. Location matters. Age matters. Credentials matter. Pay matters. So does the difference between a job opening next year and one appearing after your savings are gone.

That is where Stewart’s concern is strongest. It is also where the economic research gives him better material than an executive struggling to improvise a job title.

AI’s early record is mixed

We should apply the same care to the evidence arriving now.

A published study of customer-support workers by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond found that access to an AI assistant increased issues resolved per hour by about 15 percent on average. Benefits differed substantially across workers.

That demonstrates useful assistance in a particular setting. It does not establish that the employer will need more workers, or that the same result applies throughout the economy. Higher productivity and higher employment are separate outcomes.

Research by Anders Humlum and Emilie Vestergaard provides another piece. Their Danish study linked surveys covering 25,000 workers in 11 exposed occupations to administrative records through December 2024. Despite widespread adoption and changes in tasks, they found no detectable average effect on earnings or recorded working hours, ruling out effects larger than approximately 2 percent over the observation window.

Work was changing before the standard measures showed much movement. That finding also has boundaries: Denmark, selected occupations, and an early period of adoption.

The Stanford Digital Economy Lab’s updated research raises a different concern. By June 2026, employment among workers ages 22 to 25 in highly AI-exposed occupations was about 19 percent below where it would have been had it kept pace with similarly aged workers in less-exposed occupations.

That is a relative gap, not a finding that AI eliminated 19 percent of young people’s jobs. The researchers explicitly describe the patterns as descriptive rather than causal. They do not find widespread economy-wide displacement, but they do see a widening problem for younger workers, operating primarily through reduced hiring.

Those distinctions keep us from turning a concerning signal into a claim the study cannot support. They also identify a practical question: if companies reduce entry-level hiring, how will people acquire the experience needed for the jobs that remain?

That is a more immediate issue than guessing what a job advertisement might say in 2045.

The promises still need to be challenged

Historical comparisons also have a limit. If we are discussing systems that would eventually perform nearly every economically useful task at lower cost than people, that scenario deserves to be examined on its own terms. The history of tractors or telephone switches cannot establish that humans will retain plentiful, well-paid work under that assumption.

This is the legitimate force of Stewart’s conditional challenge. Executives cannot promise something approaching the replacement of human capability and then wave away the implications for human income.

But that remains a conditional scenario. It should not be silently substituted for what existing systems can do, how quickly organizations adopt them, or what the present evidence demonstrates.

Executives should have to explain which future they are promising. Are they anticipating tools that help people accomplish more? Systems that eliminate particular occupations? Something approaching a broadly automated economy? Those propositions have different implications.

And when they promise prosperity, they should explain the route from company revenue and productivity gains to household income. Who owns the productive assets? How do workers participate? What happens during the transition? What would cause the executives to reconsider their forecasts?

Those are harder questions than asking someone to name a future job. They are also harder to escape with a few confident sentences about human ingenuity.

A good joke should leave us understanding more

I did not look into this because I wanted to defend AI executives. I looked into it because I thought the segment left something consequential out.

There is evidence that technological change creates new work. There is evidence that workers can suffer lasting losses even as the economy adapts. Nobody gets to select only the part that makes their preferred conclusion sound inevitable.

Stewart has long distinguished his comedy from conventional journalism, including in his famous 2004 appearance on CNN’s Crossfire. I understand the distinction. A comedy show does not have to become a graduate seminar.

But the effect on an audience does not stop at the category assigned to the program. Pew Research Center reported that, in its 2014 survey, 12 percent of online American adults said they had received political or government news from The Daily Show during the previous week. The median age of those news consumers was 36, substantially younger than the news audiences of conventional television programs. Those historical figures document the show’s informational role and its appeal to a younger audience.

People can come for a laugh and leave with an understanding of the world. When a segment arranges factual claims into an argument about our economic future, that understanding matters.

Stewart could have exposed the distance between the executives’ promises and their evidence while acknowledging that uncertainty about future occupations is a normal feature of technological change. He could have made the more difficult point, too: an economy’s capacity to create new work does not tell a displaced worker whether it will help them.

That would have been a sharper argument, a fairer one, and a more useful warning.

I laughed at the clip. I think that framing was heavy-handed, unfair to the evidence, and beneath the standard he has shown he can meet. He could have done better.

Commentary reflects the author’s interpretation of the evidence. AI assisted the research and drafting. Sources are linked throughout. Read our AI disclosure.

← More commentary