We need to talk about jobs

Widespread disruption from automation is barreling toward the U.S., and everyone is hoping for the best—or too busy trying to decide which AI tool to buy

There’s a scene in “The Big Short,” where Dr. Michael Burry (played by Christian Bale), the eccentric hedge fund manager of Scion Capital, confronts the mounting pressure from his investors. Burry meticulously analyzed thousands of individual mortgages and after realizing that the industry was nearing a critical failure point, made a massive bet against the U.S. housing market.

It was a move deemed to be ludicrous by Wall Street at the time (around 2005-2007). Burry had to pay exorbitant fees to maintain his short positions while he waited for the market to realize what he had seen (a delay that his investors were furious about). During a heated exchange with one of his top investors, Burry defends himself saying, “I may have been early, but I’m not wrong.”

The months leading up to the collapse of the U.S. housing market may be a morose analogy for the U.S. labor market, and yet I worry it’s an apt comparison given the scale and impact that AI-driven automation will have on jobs, and by extension, on the education system whose purpose is in large part to prepare students for those jobs.

I’ve met with hundreds of superintendents, foundation leaders, and policymakers over the past twelve months, and if it’s an introductory conversation, it usually starts off with the assumption that the priority is figuring out how to build and deploy AI tools in education.

That isn’t it.

Focusing on tools misses the forest for the trees. I feel like Dr. Michael Burry, shouting from the rooftops about what I know to be a certainty: We are about to experience a hurricane. The skies are deceptively sunny, the air feels calm. But its always calm before the storm.

The Big Short is my second favorite movie about the 2008 financial crisis. My favorite is Margin Call. While the movie was widely praised, it is best known for one scene, where John Tuld, CEO of a fictional Goldman Sachs is briefed on the findings of two analysts who stumbled upon the same data that inspired Dr. Burry’s prescient short of the housing market. Even if you don’t have time for a feature length film, it’s worth taking 10 minutes to watch the scene—it’s is a masterpiece of modern cinema. In it, Tuld (played by Jeremy Irons), quickly comes to terms with the magnitude of what’s about to happen:

“Do you all care to know why I’m in this chair with you all? Why I earn the big bucks? I’m here for one reason and one reason alone. I’m here to guess what what the music might do a week, a month, a year from now. That’s it. Nothing more. And standing here tonight, I’m afraid that I don’t hear a thing…

This is it. I’m telling you, this is it.”

Today, this is the wave of disruption to jobs by new forms of automation that are already well underway, and taking hold at unprecedented speed.

Stumbling upon the challenge of our lifetimes

The impact of AI on jobs has been an obsession of mine since 2018. I had left politics and Washington, D.C. for the perennial fog of San Francisco to work in the AI space (initially in policy roles, and then in social impact). I had been hired by Nicolas Economou, CEO of H5, to stand up a corporate social responsibility initiative at his company. H5 was founded in 2002 and pioneered the use of AI in the law. Nicolas was a visionary, convinced that AI was on the brink of dramatically disrupting society. He brought me on to advance what he saw as two critical priorities: establishing standards for the use of AI, and building AI literacy amongst judges and legal practitioners.

I come from a family of educators, and by accident had discovered that the public schools in Akron, Ohio, where my mom still teaches today, didn’t have any curriculum that addressed AI or the future of work. I was astounded. Brookings had just published a report detailing the projected geographic impact of AI on jobs, and Akron, OH was in the top 20 cities in the country most at risk.

Even in 2018, “AI,” “the fourth industrial revolution,” and “the future of work” were phrases that had made their way into the vernacular of business and world leaders. The World Economic Forum two years earlier chose “Mastering the Fourth Industrial Revolution” as the theme of its 2016 annual meeting in Davos. If anyone should be talking about “the future of work” and the “Fourth Industrial Revolution,” its the future workers—that is, the students who were literally about to make one of the most consequential choices of their lives.

I’ll save the long version of the story of how I founded aiEDU, but the short of it is that after scouring the AI community for any examples of curriculum or instruction about AI for high schoolers, I discovered that Akron Public Schools wasn’t the exception, but the rule. In fact, back in 2018 there were virtually no schools teaching students about AI outside of the odd computer science class or summer camp.

Today, despite the fact that we are 2.5 years out from the release of ChatGPT and the ushering in of the so-called “age of AI",” precious few resources are being directed to addressing the need for urgent, system-wide investment in helping schools build the capacity to understand and prepare students for this new era. Instead, everyone seems to be focused on building tools and widgets.

Go back to the mid-1970s, when basic calculators were becoming widespread thanks to the rapidly falling cost of integrated circuits. In hindsight, you wouldn’t have focused on ‘calculator literacy’—you’d have focused on preparing people to navigate a technology revolution that would reshape the workforce and society.

I’m obviously not the only person warning about the risk AI poses to jobs, but having met with dozens of CEOs from the world’s largest companies and philanthropies over the past year, I think it’s fair to say that the “jobs issue” is being subordinated to a focus on how to “harness AI to do XYZ.” We’re in a race to the bottom, and everyone is putting blind faith in the story we are all telling ourselves: that AI is going to solve all of our problems for us. It’s going to create new jobs we can’t even imagine.

AI is an amazing technology, and we’re no doubt in store for some very exciting developments in the coming years, but it’s lazy, convenient and folly to rely on the hope that AI will solve the very problems it is going to create.

Smoke signals becoming brush fires

Data, and more importantly, a growing array of real-world examples of AI-enabled job displacement, are brining to life predictions made more than a decade ago.

Worsening job market for new grads

A recent article in The Atlantic paints a sobering picture. The New York Fed is warning that labor conditions for recent college graduates have “deteriorated noticeably.” The unemployment rate for new grads is unusually high at 5.8 percent, with graduates from elite M.B.A. programs struggling to find work. We’re also seeing a surge in law school applications, which could be a sign that young people are using graduate school as a fall back amidst a challenging job market.

This might come as a surprise, given the U.S. unemployment rate is still near historic lows. There’s a lot to be said about why this is the case, but the TL;DR is that the labor force participation rate, which establishes the denominator for the unemployment rate, is hovering around 62.5%, down from 67.2% at the beginning of the 21st century.

Once you zoom in beyond the macroeconomic picture, you also start to see cracks in the facade. The number of openings for software development and IT operations jobs is down significantly. The share of jobs posted on Indeed in software programming has fallen by more than half.

The result is the largest gap between the unemployment rate between young college graduates and the overall labor force.

As The Atlantic puts it:

“the strong interpretation of this graph is that it’s exactly what one would expect to see if firms replaced young workers with machines. As law firms leaned on AI for more paralegal work, and consulting firms realized that five 22-year-olds with ChatGPT could do the work of 20 recent grads, and tech firms turned over their software programming to a handful of superstars working with AI co-pilots, the entry level of America’s white-collar economy would contract.”

AI-driven layoffs

There are other interpretations of the job market for new grads,so let’s dig into some of the qualitative and anecdotal data to paint a picture of the early impacts that LLMs and generative AI are having on the labor market:

Intuit

The Tax-preparation software company reported last year that it plans to lay off 10% of its workforce, about 1,800 employees, to free up cash flow in what has been described as an AI pivot to grow “technology teams and capabilities in strategic locations.”

CEO Sasan Goodarzi emphasized Intuit’s focus on generative AI in a letter to employees, and committed to hiring an equivalent number of employees in “our most critical areas to support our customers and drive growth.”

Cisco

Cisco announced plans last year to lay off 7% of its workforce, roughly 5,900 employees, as part of a cost-management effort amid declining demand, as well as a stronger focus on the company’s AI strategy. This followed a previous reduction in force by 4,000 employees earlier in 2024.

IBM

IBM told the WSJ last week that they tapped AI to automate the work of “several hundred human resources employees.” Earlier this year, IBM CTO Ji-eun Lee said that their AskHR agent had automated 94% of simple, routine human resources tasks; and their AskIT agent reduced the number of calls and chats for the IT team by 70%.

Over the past few years, IBM’s AI division has grown to become a $6 billion business, with a similar suite of offerings to Amazon, Salesforce, and Microsoft.

Activision Blizzard

The game development industry was recently the focus of a WIRED article looking into some of the ways AI is displacing jobs after a survey found that half of game developers say their workplace is using AI, and four out of five have concerns about its use. WIRED reports:

“Managers at video game companies aren’t necessarily using AI to eliminate entire departments, but many are using it to cut corners, ramp up productivity, and compensate for attrition after layoffs. In other words, bosses are already using AI to replace and degrade jobs. The process just doesn’t always look like what you might imagine. It’s complex, based on opaque executive decisions, and the endgame is murky. It’s less Skynet and more of a mass effect—and it’s happening right now.”

“It’s here. It’s definitely here, right now,” says Violet, a game developer, technical artist, and a veteran of the industry who has worked on AAA games for over a decade. “I think everyone’s seen it get used, and it’s a matter of how and to what degree. The genie is out of the bottle, Pandora's box is opened.”

Last year, Activision, one of the largest AAA video game developers in the world, began pushing the use of AI to the chagrin of software engineers and artists on staff who worried about the long term implications for their own jobs. Their fears were realized only months later, when 1,900 Activision Blizzard employees were laid off. Among the hardest hit teams were 2D artists. The above image was a Christmas themed loading screen in the popular Call of Duty franchise, which appears to have been created using AI (the zombie has 6 fingers).

Duolingo

Duolingo CEO Luis von Ahn announced that his company was officially “going to be AI-first” in an all staff email on April 29. Von Ahn wrote that the company will “gradually stop using contractors to do work that AI can handle.” Brian Merchant spoke with some of those contractors, and they revealed that that this isn’t a new initiative and in fact, Duolingo has already replaced up to 100 of its workers (primarily writers and translators) with AI systems: “The translators were laid off in 2023, the writers six months ago, in October 2024.” One writer told Merchant:

“If you had asked me a year ago, I would have told you that my job would become more and more editing AI content. I did not expect to be replaced so soon.”

Duolingo has reported that the integration of generative AI has significantly accelerated their ability to build content. In a blog post, von Ahn wrote:

“Developing our first 100 courses took about 12 years, and now, in about a year, we’re able to create and launch nearly 150 new courses. This is a great example of how generative AI can directly benefit our learners, and reflects the incredible impact of our AI and automation investments, which have allowed us to scale at unprecedented speed and quality.”

Klarna

In December last year, Klarna instituted a hiring freeze as it reportedly sought to focus on AI investments. Headcount fell by 22%, mostly due to attrition, with CEO Sebastian Siemiatkowski calling for employees to turn to AI to help fill the gaps. Klarna claimed that AI could replace the work of 700 customer service agents and that it had taken over 75% of the customer chats (about 2.3M) per month, with AI bots handling questions about topics like refunds, returns, and payments in more than 35 languages. Earlier last year, Siemiatkowski claimed that AI could help the company reduce headcount by 50%.

While I was writing this piece I got a text from Peter Relan, founder of MathGPT.ai and initial investor in Discord, with a link to an article reporting that Klarna has just admitted they were overly bullish on AI capabilities, and lifted their hiring freeze to bring more human agents to improve the customer experience. I haven’t seen any data but I think it’s safe to assume there will still be a net reduction in employment at the company, even if the retraction falls short of the lofty predictions being made by their CEO.

UPS

Last month UPS announced plans to cut 20,000 jobs and shut down 73 facilities across its U.S. network in 2025. The move comes amidst revelations about the company’s potential partnership with humanoid robot company Figure AI. Figure made waves last year when they released a demo of their new speech and reasoning capabilities (the video has nearly 3M views and was a mainstay of aiEDU presentations about AI development). Coincidentally, Figure AI posted a new video in February showing a team of its robots packing parcels in a sorting facility.

UPS has not confirmed the partnership, stating that the company “regularly explores and deploys a wide range of technologies, including robotics,” but logistics has long been a frontier for deploying robotics. That brings us to…

Amazon

Last week, Amazon—which has already deployed over 750,000 robots across its fulfillment centers—announced a “fundamental leap forward in robotics” after developing a robot with a sense of touch that has the ability to handle three-quarters of the items in its global network of warehouses. As The Verge reports, “Vulcan is not Amazon’s first robot capable of picking items up, but it is the first that’s dextrous and sensitive enough to maneuver goods inside the compact, fabric-covered compartments that the company uses for storage — which are themselves already moved around warehouses by a different fleet of robots. Vulcan uses an arm that Amazon says “resembles a ruler stuck onto a hair straightener” to rearrange any items already in a compartment and add new ones, with force sensors that help it know when it makes contact with an object and how much force and speed to use to avoid causing damage. A second arm includes a suction cup to grab anything it wants to take out of the pods, with an AI-powered camera to make sure that it hasn’t picked up multiple items by mistake.”

Gartner predicts that 75% of large enterprises will integrate AI enabled robotics into their warehouse operations by 2026, driven by increased efficiency, lower labor costs, improved safety, and improved customer satisfaction. (Their report includes a footnote about a potential downside being the “potential for job displacement, as certain human-performed tasks may become automated.”)

Amazon is the world’s second largest employer, with over 1.5M people on payroll. That’s a massive number, but it’s 100,000 fewer than the 1.6M people employed in 2021. Over the same period, the company grew the number of robots deployed by more than 250,000…

IKEA

The Swedish furniture giant announced in 2023 that it has successfully re-skilled 8,500 call center workers to serve as interior design advisors thanks to the introduction of Billie, the company’s AI customer service agent inspired by their popular range of bookcases (I still have one of these in my childhood bedroom back in Akron, Ohio).

IKEA’s global people and culture manager has said that they are “committed to strengthening co-workers' employability through lifelong learning and development and reskilling,” and did not report a reduction in headcount at the company. However, Billie has reportedly handled 47% of customers’ call center queries over the past two years.

Ride-hailing

If you live in San Francisco, Los Angeles, Austin, Las Vegas, Phoenix, San Diego, Dallas, Miami, or Atlanta, its likely that you’ve seen Waymo’s iconic self-driving Jaguar I-PACE vehicles. As a local San Franciscan, I can attest to Waymo becoming the dominant mode of transportation for my “techie” friends. Estimates of Waymo’s market share are hard to pin down, but the company passed 537,000 rides per month in December 2024 in SF and LA alone, with monthly growth between 16K-66K rides, which translates to between 3-18% of market share per year.

If Waymo sustains the 18% trajectory, it would put it on par to the growth we saw in ride-hailing between 2013 and 2018, a period where the number of taxi drivers doubled.

There are a lot of factors in play if one aims to forecast the impact of autonomous vehicles on the ~2M ride-hailing and taxi drivers in the U.S. Regulation, cost, safety, and public perception are all known unknowns. But I think it’s safe to say that driving for Uber and Lyft are unlikely to be sustainable long-term career pathways. Even in the conservative projection above from Ethan Heppner suggests a retraction of opportunities. This might just mean daily revenue for drivers decreases as they compete with autonomous vehicles, moving the gig work closer to minimum wage levels.

The freelance job market

Digital transformation and change management takes time and political capital. Even if AI is today capable of overhauling the way companies do business, there are significant barriers to wide scale implementation at large organizations (i.e. bureaucracy and a low supply of AI knowledge). The freelance job market is far more elastic, on the other hand.

What we’re seeing is about a 20% decrease in freelance jobs that require basic writing, coding or translation across platforms like Upwork and Fiverr.

It’s worth noting that 64 million Americans do some kind of freelance work each year.

Labor economists warned us about this

Driving my concern about jobs was mounting research from economists about the massive potential for AI to disrupt the labor market. As early as 2013, an Oxford University study that evaluated the potential for AI (primarily machine-learning at the time) across 702 different job categories in the U.S. concluded that fully 47% could be automated “over the next decade or two.” A working paper from the OECD in 2018 found that 14% of jobs have at least a 70% chance of automation, with a further 32 percent with a probability between 50-70%. At current employment rates, that implicated 210M jobs across the 32 countries in the study.

Robert Atkinson, president of the Information Technology & Innovation Foundation, published a rather glib op-ed in 2022 titled “Oops: The Predicted 47 Percent of Job Loss From AI Didn’t Happen.” Atkinson admonishing the 2013 Oxford study as “alarmist” and claiming it “significantly overstated the share of jobs at risk.”

Just two months later, OpenAI released ChatGPT to the world.

Betting against the steady advance of technology is perennially risky, and I take umbrage at the misrepresentation of the Oxford study’s prediction—the authors did not claim that 47% of jobs would disappear in a decade. Here’s what they said (emphasis mine):

“47 percent of total US employment is in the high risk category, meaning that associated occupations are potentially automatable over some unspecified number of years, perhaps a decade or two.”

Semantics aside, I think even Mr. Atkinson would admit that given the pace of development over the past two years another decade of progress in the capabilities of generative AI, it is highly plausible that a large share of U.S. employment is at risk of automation.

My own cursory review of total U.S. employment suggests that the Oxford researchers were on to something, even if they could not have predicted precisely how AI capabilities would accelerate.

Since the release of ChatGPT ushered LLMs and generative AI into the public zeitgeist, the U.S. economy has created 8.1M new jobs. With the unemployment hovering near 4% amidst these new AI capabilities, the consensus seems to be that AI is creating more new jobs to balance out the jobs it displaces. I’m not so sure.

From brush fires to forest fires

I’m known to belabor analogies and will do so here. Current iterations of AI are the sparks and matches and the low hanging fruit for AI implementation (software engineering, customer service, warehouse robotics, autonomous transportation etc.) are the brush. Sure, we see smoke, and there are some small fires that could get bigger, but to most observers, all they see is a vast forest, and while it might be a tad dry, everything seems just fine.

I’m convinced that we are about to see what happens when you pour an Olympic pool’s worth of gasoline on those brush fires—a recession.

Before I explain, let’s get one thing out of the way, no one knows how long we have until the next recession. J.P. Morgan last month upped its estimation of recession risk in 2025 to 60%. Goldman Sachs’ chief economist published a podcast last week where he estimated a 45% chance of recession in the next 12 months. HSBC predicts a 40% chance.

Regardless of whether we are 12, 18, or 24 months away from a downturn, modern economic theory dictates that sooner or later, we are going to find ourselves in a recession. Cue the gasoline.

Historically, job displacement due to automation doesn’t happen slowly; it comes to a head all at once. This 2018 analysis by Nir Jaimovich and Henry Siu for the National Bureau of Economic Research (NBER) found:

“the disappearance of per capita employment in routine occupations associated with job polarization is not simply a gradual phenomenon: the loss is concentrated in economic downturns. Specifically, 88% of the job loss in these occupations since the mid-1980s occurs within a 12 month window of NBER dated recessions…Essentially all of the contraction in per capita aggregate employment during NBER dated recessions can be attributed to recessions in these middle-skill, routine occupations. Jobless recoveries are observed only in these disappearing, middle-skill jobs.”

Put simply, over the last 50 years, a combination of mechanistic automation, off-shoring, and other factors drove a polarization of the job market, pushing out workers in routine, but well-paying, middle-income jobs toward either low-paying but hard-to-automate service jobs or high-paying knowledge work. This is the ‘death of middle class jobs’ that we hear about so much (especially for me, as a native Ohioan who grew up in one of the hardest-hit regions in the U.S.). This intuitively makes sense. It’s complicated and painful to use technology to replace people. You have to restructure organizational charts, implement unpopular change management, etc.

But in a recession, everything changes. Imagine you are the CEO of a large company with a gross margin of 20%, and you learn that profits are on track to be down by $1Bn. You essentially have two options to addressing the deficit: increase sales by $5Bn, or cut costs by $1Bn. Furthermore, you can depreciate your capital expenditures (CapEx) to reduce your taxable income—a benefit that isn’t available to companies that are considering investing in their workforce.

When (not if) the next recession comes, CEOs are going to have 17+ years of technology overhang, and a powerful slate of generative AI capabilities, at their disposal. The NBER research makes it very clear what happens next: companies implement reductions in force, and turn to the latest technology to bridge the operational gap.

But isn’t AI going to create new jobs?

In defense of those who are taking a ‘glass half full’ mentality to AI and automation, it is less a question of if AI will replace jobs, but rather a discussion of how steep the curve of displacement is. After all, we’ve experienced myriad instances of automating technologies displacing jobs, but if you zoom out sufficiently far, the story of technology innovation is a story of job growth, not job contraction. Historically, new technologies that displace jobs have unfolded over a period of decades, giving the labor market ample space for demands for new expertise to create new opportunities for displaced workers. This was the case for the printing press, electricity, automobiles, computers, and the internet.

So, the common refrain goes, AI may indeed destroy some jobs, but it will also create new jobs—jobs we can’t even imagine today.

I’m not so sure.

Past technologies that automated tasks and jobs indeed created huge numbers of new jobs, often in categories that were directly related to the technology itself. Microsoft Excel, for instance, significantly increased the productivity of accountants, but the broader computer revolution created entire new industries. Suddenly, companies needed to hire IT administrators, web managers, technology procurement managers, and cybersecurity specialists. New multi-billion dollar retail chains like RadioShack and Circuit City (however long they lasted) flourished, built around selling and supporting these new technologies, alongside the complex new supply chains and logistics networks required to sustain them. Streaming video disrupted and eventually destroyed the physical VHS and DVD industry (sorry Blockbuster!), but the technology, especially when converged with social media and YouTube, massively increased the quantity of video content produced and the roles associated with its creation, distribution, and monetization.

I'd posit that before the personal computer revolution, it would have been fairly obvious that a significant number of new jobs would be required if every single worker and household eventually had at least one computer. We would need people to build them, sell them, install them, network them, repair them, and develop software for them.

And yet, with AI, I haven't heard a single compelling case for what all these new, large-scale job categories might be. A year ago, many claimed these jobs would be "prompt engineers." I had my doubts then; it seemed obvious that AI would improve to the point where constructing sophisticated prompts would no longer be a specialized, standalone necessity for most users. My skepticism appears to have been warranted. Last week, Fast Company published a story detailing the disappearance of prompt engineering roles:

"Just two years ago, prompt engineering was hailed as a hot new job in tech. Now it has all but disappeared. At the beginning of the corporate AI boom, some companies sought out large language model (LLM) translators—prompt engineers who specialized in crafting the most effective questions to ask internal AIs, ensuring optimal and efficient outputs. Today, strong AI prompting is simply an expected skill, not a stand-alone role. Some companies are even using AI to generate the best prompts for their own AI systems. The decline of prompt engineering serves as a cautionary tale for the AI job market. The flashy, niche roles that emerged with ChatGPT’s rise may prove to be short-lived. While AI is reshaping roles across industries, it may not be creating entirely new ones."

What if AI creates lots of bad jobs?

Even if AI creates more jobs than it displaces, there’s no guarantee those jobs will be better.

The loom, which kicked off the first Industrial Revolution, sparked concern among many, perhaps most notably Karl Marx. Marx and others predicted that automation would result in mass unemployment. But that’s not what happened. By the end of the 19th century, there were four times as many factory weavers as in 1830. The loom lowerd the price of cloth, which generated a massive increase in demand, and more jobs for weavers.

Here’s the thing: Working in a textile factory sucks. Conditions in the 19th century were horrific (read about the Triangle Fire). These new jobs, while plentiful, where characterized by grueling monotony, dangerously unsafe environments, and profound exploitation. Workers, including vast numbers of women and young children, toiled for 12 to 16 hours a day, six days a week, for pitiful subsistence wages. The factories themselves were often poorly lit and terribly ventilated. Accidents were horrifyingly common.

The story about the Industrial Revolution usually focuses on the tremendous GDP gains that technology unlocked, but we forget that the jobs lost to these new factories were a dream by comparison. Consider the life of cottage weaver before their livelihoods were destroyed: you worked from home, owned your own loom, picked your clients, and could set your own schedule. That isn’t to suggest cottage weavers didn’t work hard, but they captured the value of their own labor. It was skilled work, wrapped in a measure of autonomy and dignity that disappeared the moment spinning jennies and power looms arrived.

It’s all about expertise

Innovation doesn’t necessarily lead to bad outcomes. Quite the opposite. Technology has largely unleashed previously unimaginable levels of productivity and automated mundane aspects of work. We used to employ people who worked around the clock as human calculators, stenographers, and bookkeepers—to say nothing of the millions of farmers who ceased toiling in their fields thanks to mechanization.

Expertise is key to understanding why some technologies augment humans, while others replace us. This is the focus of research by David Autor, a labor economist at MIT who has been one of the leading scholars working on the question of how AI is going to impact jobs.

Autor has dozens of published articles, but his latest research delves into the connection between expertise and jobs in a work titled “New Frontiers: The Origins and Content of New Work, 1940-2018.”

Autor and his co-authors estimate that roughly 60 percent of the jobs Americans hold today are carried out under titles that did not exist in 1940. That startling reality does two things at once. First, it punctures the “fixed-pie” fear that a wave of automation simply leaves fewer slices for humans; we have been inventing entirely new pies for eight decades. Second, it reframes the AI debate: the real question is not whether we will have work, but what kind of new work we will create next—and for whom.

Autor shows that the birthplace of new work has migrated. Between 1940 and 1980, fresh job titles clustered in production and clerical occupations—the solid, middle-skill center of the labor market. Since 1980, however, most new titles have emerged at the high-skill professional end and, secondarily, in low-paid personal services, leaving the middle dramatically thinner. That pattern lines up with the “barbell” the article traced earlier: a swelling cohort of well-paid knowledge jobs at one end, growing in-person service work at the other, and a hollowing centre.

Using a patent-level analysis, Autor et al. distinguish augmentation innovations (which complement what people produce) from automation innovations (which substitute for what people do). Only augmentation patents reliably predict the appearance of new titles; automation patents do not. In practice, that means ultrasound machines created demand for sonographers, while spreadsheet software erased legions of bookkeepers.

Those two innovation streams don’t just differ in flavor; they move the labor market in opposite directions. Occupations that are heavily exposed to augmentation patents grow in employment and wage-bill; occupations deeply exposed to automation shrink. In short, augmentation creates new demand for new types of labor, while automation erodes it.

In a lecture about his research, Autor reminds us that the truly scarce asset is domain-specific expertise—knowledge that lets someone accomplish a valuable objective. AI can either lift that expertise, much like a pneumatic hammer amplifies a roofer’s skill, or commodify it, turning seasoned London cabbies into drivers who “feed the monkey” while Waze does the real navigation.

The same large language models that draft amicus briefs in seconds can also tutor a novice paralegal on the logic behind each clause. Whether the technology displaces or democratizes expertise is not preordained; it will be shaped by incentives, regulation, and, crucially, education.

The imperative for K-12 education

If expertise is the scarce asset, then K-12’s core mission cannot stop at “AI literacy.” Treating AI like a stand-alone subject would be as myopic as mandating “mobile-phone literacy” in 2007. By the time statewide curriculum or standards are in place, the average 8th grader, an AI native who has spent half her life interacting with LLMs, will be running circles around us adults. We should expect younger generations will be teaching adults how to use AI, not the reverse.

The real moon-shot is whole-school transformation that systematically cultivates durable skills (or 21st century skills) like critical thinking and problem solving—while doubling down on foundational knowledge in math, English, science, social studies, and computer science. AI cheapens access to information and content production, but raises the premium on knowing which questions to ask and how to judge the answers. It’s back to the future, folks. Mastery of the basics will be essential knowledge to ensure that students aren’t the modern day, chatbot manager equivalent of factory workers.

The good news is that instruction that builds those durable skills can be integrated into core subjects and also helps to address one of the biggest challenges in today's education system: student's feelings that what they are learning isn't relevant or engaging.

It looks like a middle school math class using a section about fractions to teach students about how language models blend the probabilities from thousands of neurons to predict the next word, and then having students run a simulation, adjusting the weights like fraction coefficients, to see how tiny changes can tilt the outputs.

It looks like high school English class project where students build a role playing game based on Of Mice and Men, and then giving a TED talk in the persona of the character they played.

I certainly have a lot more to say about the future of K-12 education, but I’ll save it for a future post.

A Final Note of (Conditional) Optimism

Michael Burry was early but not wrong; the same may be true of today’s automation Cassandras. Yet Autor offers a practical goal to anchor the urgent work that must be done—channel AI toward amplification, not substitution, and the next tranche of new, fulfilling, better-paid work will arrive right on schedule.

We just need to prepare people to seize it, and we have our work cut out for us.