Will AI Create or Destroy Jobs? Myth, Reality, and a Clearer Way to Tell

You’ve heard the myth. World-changing technologies, optimists say, may destroy some jobs in the short run, but always end up creating more in the long run. Light bulbs killed candle-makers but built an entire electrical industry. Steam engines displaced canal workers and produced the modern factory economy. The internet hollowed out classifieds and travel agencies, then spawned millions of jobs that didn’t exist before. Productivity rises, prices fall, demand expands, and net employment ends up higher than it started. Always.

Then AI shows up, and the question everyone starts asking is “is this time different?” Even careful thinkers ask it that way. I’ve caught myself asking it that way.

That’s the wrong question.

The premise is that there’s a stable pattern called “what technology does to jobs,” and the only thing in dispute is whether AI breaks the pattern. Problem is, the pattern is an illusion.

Sometimes it’s happened. Other times it hasn’t. The job implications people point to are real for some technologies in some periods. Wrong for other technologies in other periods. Most importantly, the difference isn’t in the technology itself. It’s in the choices people and businesses make about how to use it. So “is AI different?” can’t be answered as a yes-or-no, because AI itself isn’t doing one thing.

Technologies aren't job-creating or job-destroying. Technologies are neutral, and it's the choices we (humans) make that give them job-creating or job-destroying potential. Neither fate is inevitable. Said differently, specific people, making specific choices, give AI its job-creating or job-destroying effects. Not some inherent property of the technology. Not “society” in some existential sense.

Start with the case the optimists love. ATMs were supposed to wipe out bank tellers. Instead, tellers grew. ATMs made it cheaper to run a branch, so banks opened more branches and hired more tellers. That story is true, and for two decades it was the canonical proof that automation creates jobs through the back door. Then it stopped being true.

Bank teller employment has fallen nearly 30% since 2010. Job postings are down about two-thirds.¹ The Bureau of Labor Statistics projects another 13% decline through 2034.² Branches are closing.³ When ATMs spread widely in the 1980s and 1990s, the number of tellers required to operate a typical urban branch fell from about 20 to 13, but banks responded by opening many more branches, and total teller employment rose with them.⁴ That dynamic held for two decades. Then mobile banking arrived, branch traffic collapsed, and the race tipped the other way.⁵ Same technology, same industry, job creation for twenty years and job destruction for the last fifteen. The optimists’ favorite example has, quietly, become the opposite.

This isn’t a one-off. James Bessen at Boston University studied two centuries of US data on cotton textiles, primary steel, and automotive manufacturing. All three followed the same inverted-U pattern: productivity grew the whole time, employment also grew for many decades, then peaked and declined. In 1958, the US broadwoven textile industry employed over 300,000 production workers and the primary steel industry employed over 500,000. By 2011, broadwoven textiles employed only 16,000 and steel only 100,000.⁶ 

Productivity didn’t reverse. It just went from creating jobs to destroying them at different points in history (same as banking). Like technology, “productivity” itself doesn’t inherently create or destroy jobs either.

There’s an argument from economics called Jevons Paradox, originally from William Stanley Jevons’ 1865 book The Coal Question, where efficiency gains in steam engines expanded total demand for coal rather than shrinking it.⁷ Jevons is frequently cited by people who believe net job creation is inevitable. Problem is, Jevon’s argument was about resource consumption, not jobs. Jevons himself never wrote a word about jobs. People borrow his name for a labor claim he never made.

So, the universal optimist claim isn’t quite right. Neither is the universal pessimist claim. The real question is the conditional one. Under what circumstances does the use of a technology create jobs? Under what circumstances does it destroy them?

One technology, opposite job impacts

Look at ridesharing.

Uber initially enabled a population of people who couldn’t have been paid drivers before to become paid drivers for the first time. The barrier was on the supply side. Pre-Uber, being a paid driver required either buying a taxi medallion (which peaked at over a million dollars in New York in 2013)⁹ or getting hired by a dispatch company with a regulated fleet.

Post-Uber, you needed a car and a phone. The Census Bureau’s own economist on the project called this “a prime example of technology lowering the barriers to entry and affecting how individuals earn money in the labor market.”¹⁰ Self-employed drivers in the taxi and limousine industry grew from about 170,000 in 2010 to over 1.3 million by 2019, with most of the new entrants being people who would not otherwise have entered the industry: younger, more often female, and more often combining driving with a separate wage-and-salary job.¹¹ The displaced wage-and-salary taxi-driver population was a small fraction of that.¹² Net result: more than a million new participants entered the paid driving market because the barrier to being a seller collapsed.

Waymo is different. Sure, it’s ridesharing, but Waymo doesn’t enable a new population of drivers. Instead, it removes the driver entirely from a market Uber already created and expanded. Waymo launched the first commercial robotaxi service in metro Phoenix in December 2018,¹³ and as of early 2026 operates fully autonomous public service in ten US metros, with about 500,000 paid trips a week and plans to reach a million weekly trips by year-end.¹⁴ The riders are the same riders. The trips are the same trips. The labor input is the change, and the change is downward. There is no new population of nonparticipants being pulled into the market. The barrier that fell with Uber (you can be a paid driver) is irrelevant in a market where there’s no driver.

Same industry. Two strategic moves a decade apart, with opposite labor outcomes. The first move created more than a million new jobs by enabling people who couldn’t have participated. If autonomous vehicles continue to scale, the second move is on track to remove most of them, plus the residual taxi drivers, by serving the same customers with less labor per ride.

The chart above is illustrative only, but it mimics what happened in the case of bank tellers and ATMs. Whether the reversal plays out depends on how fast autonomous vehicles scale, how regulators respond, and what fraction of trips remain better served by human drivers. The point is structural, not predictive. The same industry can host both a market-creating use of technology and a sustaining substitution, in sequence, with opposite effects on jobs.

Anyone watching a Waymo drive past in San Francisco is looking at the same broad industry that produced Uber, but a fundamentally different strategic choice with the opposite labor effect. Lumping them together as “AI in mobility” or “ridesharing automation” hides what’s actually happening.

The distinction that matters

In The Innovator’s Solution, Clayton Christensen drew a line that most people skim past.⁸ The line is where the labor story lives. It’s worth slowing down for.

A use of technology is market-creating when it enables a population to participate in the market for the first time, as buyers or sellers. It does so by lowering one or more key barriers (cost, skill, expertise, location, capital, time, physical access) that were previously keeping that population out. Once the barrier falls, they enter. Uber did this on the supply side of paid driving. Etsy did it for crafters. YouTube did it for video creators. Substack did it for writers.

A sustaining use, by contrast, targets existing customers in existing markets and delivers better performance or greater efficiency on dimensions those customers already value. Sustaining uses are often cheaper and better at the same time. The efficiency gain comes from reducing the human input required per unit of output. Waymo is doing this to ridesharing. Self-checkout did it to grocery. Streaming did it to video rental.

The distinction is that “cheaper for the people who were already there” doesn’t qualify as market-creating, even if a few marginal newcomers come along. The primary mechanism of market creation is the entry of a population that couldn’t participate before.

The travel industry makes this concrete. Expedia and Kayak served existing travelers a more efficient way to do something they were already doing. The people booking on Expedia were the same people who used to book through travel agents. The barrier removed was the friction of going through an agent, not the barrier to traveling at all. Travel agent employment fell from about 105,000 in 2002 to 65,700 in 2024.¹⁵ That’s sustaining, and the labor effect is what sustaining does.

The reason the labor effects diverge between market-creating and sustaining strategies is structural. Market-creation adds a population that wasn’t participating before. That population is additive almost by definition, because previously-excluded participants don’t displace anyone. They weren’t there. Sustaining strategies serve populations that were already there with less labor per unit, which removes jobs almost by design, because that’s where the efficiency gain comes from.

If you want to know whether an innovation’s strategy will be job-creating or job-destroying, three questions can help:

Question 1: Is there a population that could not participate in this market before, who now can because a cost, skill or physical barrier has fallen?

Yes → market-creating signal

No → sustaining signal

Question 2: Is the customer base for this new innovaiton the same customer base that existed before, or substantially the same?

Yes (same customer base) → sustaining signal

No (different or expanded customer base) → market-creating signal

Question 3: Where does the value come from: from enabling new participants, or from doing the existing work with fewer people?

Enabling new participants → market-creating signal

Doing existing work with fewer people → sustaining signal

A more thorough explanation of this dynamic can be found in Christensen’s book The Innovator’s Solution⁸ (if you haven’t read it, you should). Both strategies can be legitimate in different contexts. They just tend to produce opposite labor outcomes, by mechanism, not by accident.

Where is AI today?

So which way is AI going? You can answer that case by case.

Customer service automation, which is everywhere right now, isn’t reaching customers who couldn’t get service before. It’s serving the same customers with fewer human agents. The barrier that fell was on the company’s costs, not on customer access.

That’s sustaining, and the labor effect follows.

Content production at the volumes companies are now generating with AI is the same story. The audience is the same audience. The output is replacing what used to come from freelance writers. Code generation aimed at reducing entry-level engineering headcount is again the same. The engineers being displaced were already employed. The code being written was already going to get written. Same customers. Same work. Fewer humans.

Now look at a different kind of case. People who couldn’t afford a lawyer can now get legal advice through AI. People who couldn’t access a tutor can now get one. Translation, design, analysis, things that used to require expensive expertise or were just out of reach for most people, are showing up in places they weren’t before. Solo entrepreneurs are running companies that would have required ten people five years ago. These aren’t existing customers getting served more efficiently. They’re new participants entering markets they were excluded from. In each case, the population that’s now participating wasn’t there before.

Same models. Same APIs. Same underlying capability. The strategy gives it its character.

\Most AI uses right now are in the first bucket, not the second. That’s not surprising. The cost-cutting move always arrives first with a general-purpose technology, before the growth move catches up. Steam engines spent decades cutting labor inputs in textiles and mining before railways and steamships pulled enormous new populations into paid work. AI may follow the same arc. It also may not. The arc isn’t automatic. It depends on what the people building, funding, and using AI choose to build.

Net job outcomes follow from those choices. Founders deciding what to build. Product managers deciding what to ship. Investors deciding what to fund. Teachers deciding how to teach. Writers deciding how to write. Anyone using AI is making a call, even if they don’t think of it that way. If you want this question to have a different answer, the answer will come from how each of us uses it. So ask yourself this. How do you plan to spend the next ten years living and working with AI? Would you rather expand the pie, or cut it into ever-smaller pieces? Do you want to create new markets, or cut costs in existing ones?

That’s a better question.

Endnotes

1  Burning Glass Institute, “The Case of the Vanishing Teller: How Banking’s Entry Level Jobs Are Transforming” (2025), reporting a roughly 30% decline in teller employment since 2010 and a near two-thirds drop in job postings. burningglassinstitute.org/bginsights/the-case-of-the-vanishing-teller

2  U.S. Bureau of Labor Statistics, Occupational Outlook Handbook, “Tellers” (2024 edition, projecting 2024–2034). Tellers held about 347,400 jobs in 2024; employment is projected to decline 13% from 2024 to 2034. bls.gov/ooh/office-and-administrative-support/tellers.htm

3  FDIC, Summary of Deposits, annual branch counts. US bank branches peaked near 99,000 in 2012 and have declined since. fdic.gov/sod

4  James Bessen, “Toil and Technology,” IMF Finance & Development (March 2015): “Thanks to the ATM, the number of tellers required to operate a branch office in the average urban market fell from 20 to 13 between 1988 and 2004. But banks responded by opening more branches to compete for greater market share. Bank branches in urban areas increased 43 percent.” imf.org/external/pubs/ft/fandd/2015/03/bessen.htm

5  U.S. Bureau of Labor Statistics, Occupational Outlook Handbook, “Tellers”: “the number of bank branches has been in decline due to technological change. As more people use online banking tools, such as mobile check deposits, fewer bank customers will visit the teller window.”

6  James Bessen, “Automation and Jobs: When Technology Boosts Employment,” Economic Policy 34, no. 100 (October 2019): 589–626. Bessen states: “In 1958, the US broadwoven textile industry employed over 300 thousand production workers and the primary steel industry employed over 500 thousand. By 2011, broadwoven textiles employed only 16 thousand and steel employed only 100 thousand production workers.” Working-paper version: bu.edu/law/files/2017/04/autogro.pdf

7  William Stanley Jevons, The Coal Question: An Inquiry Concerning the Progress of the Nation, and the Probable Exhaustion of Our Coal Mines (London: Macmillan, 1865). Jevons observed that improvements in steam-engine efficiency led to higher, not lower, total coal consumption.

8  Clayton M. Christensen and Michael E. Raynor, The Innovator’s Solution: Creating and Sustaining Successful Growth (Boston: Harvard Business School Press, 2003). Chapter 2 develops the distinction between sustaining innovation, low-end disruption, and new-market disruption, and offers diagnostic questions for evaluating which category a given innovation falls into.

9  AEI, “NYC Taxi Medallion Prices Have Flatlined” (analyzing NYC Taxi & Limousine Commission data), reporting individual medallion prices peaking at $1,051,000 in June 2013. Subsequent declines documented in Bloomberg Businessweek and academic analyses including the AEA paper “A Tale of Two Cities: An Examination of Medallion Prices in New York and Chicago.” aei.org/carpe-diem/nyc-taxi-medallion-prices-have-flatlined

10  James Spletzer, principal economist, U.S. Census Bureau Center for Economic Studies, quoted in “What May Be Driving Growth in the ‘Gig Economy?’” (Census.gov, 2018). census.gov/library/stories/2018/08/gig-economy.html

11  Katharine G. Abraham, John C. Haltiwanger, Claire Y. Hou, Kristin Sandusky, and James R. Spletzer, “Driving the Gig Economy,” NBER Working Paper 32766 (2024). Findings on demographic shifts among new entrants come from the paper’s analysis of Census nonemployer sole proprietor records linked to administrative wage data. nber.org/papers/w32766

12  U.S. Bureau of Labor Statistics, Occupational Employment Statistics, “Taxi Drivers and Chauffeurs” historical series. Wage-and-salary taxi driver and chauffeur employment declined modestly over the same period, dwarfed by the more than one-million-person increase in self-employed drivers documented in note 10.

13  Waymo, “Riding with Waymo One today” (press release, December 5, 2018). Coverage at TechCrunch, December 5, 2018: techcrunch.com/2018/12/05/waymo-launches-self-driving-car-service-waymo-one

14  Waymo, “Ready to Ride: Dallas, Houston, San Antonio, and Orlando” (company blog, February 24, 2026), confirming public service in ten US metros: Phoenix, San Francisco Bay Area, Los Angeles, Atlanta, Austin, Miami, Dallas, Houston, San Antonio, and Orlando. Atlanta and Austin are accessed through partnership with Uber. waymo.com/blog/2026/02/dallas-houston-san-antonio-orlando. Ridership and growth target reported by TechCrunch, “Waymo robotaxis are now operating in 10 US cities” (February 24, 2026): about 500,000 paid trips per week, with co-CEO Tekedra Mawakana stating Waymo is “on track to serve over one million rides per week by the end of this year.” techcrunch.com/2026/02/24/waymo-robotaxis-are-now-operating-in-10-us-cities

15  2002 figure from U.S. Bureau of Labor Statistics, Occupational Employment Statistics 2002, “Travel Agents” (SOC 41-3041): 104,550 employed. 2024 figure from BLS Occupational Outlook Handbook, “Travel Agents”: “Travel agents held about 65,700 jobs in 2024.” bls.gov/ooh/sales/travel-agents.htm