In September 2024, I finished a paper for Capital One examining the potential impact of artificial intelligence and automation on employment, consumer finances, and the banking system.
I’d been working on the problem since my first encounter with ChatGPT in late 2022. That experience, which I described in Twelve Seconds, changed what I thought was possible with AI.
The first thing I felt was excitement. I sensed, before I could put it into words, that the programmer’s craft was about to become commoditized. My whole career, I’d waited for the day when a misplaced semicolon stopped mattering. Knowing that kind of detail made you valuable. Applying the logic was what created value. Now the machine could manage the details, and the craft could move to the logic, the context, and the orchestration.
Ideas I’d carried around for years, I could finally build myself. For example, I modernized Horse Match, a horse marketplace, and began expanding it into horse and stable management. I’m also working on forecasting systems that select and optimize their own models. I could see an explosion of small businesses that could build their own software. I hoped healthcare improvements and drug discovery would move exponentially faster.
Immediately, the dread arrived in the form of a question I've spent years researching: what happens when machines can do much of the work people are paid to do?
Then and now, the reassuring answer has been familiar. We’ve been through technological revolutions before. The mechanization of agriculture, industrial manufacturing, computers, and the internet all disrupted employment. New industries and occupations eventually replaced the old ones. It will be fine.
But what if AI caused displacement across almost every industry, across an unusually wide range of occupations, all within a single decade?
I began researching why that might happen. Eventually, I settled on four characteristics that made this revolution fundamentally different: speed, scope, size, and the convergence of big data, computing power, and attention.
I remain hopeful, but believe that the economic consequences could be enormous, including the possibility of significant unemployment and chaos around job transitions.
How I concluded this time would be different
Speed
In 1870, almost 50% of American workers were employed in agriculture. By 1980, that figure had fallen to about 4%.
That was a massive change. It also unfolded over more than a century.
The internet moved faster. A commonly cited estimate puts its growth to 100 million users within seven years. But widespread access was only the beginning of its economic impact.
In 1999, online sales represented just 0.6% of US retail spending. Twenty years later, that figure had grown to approximately 11%. Even where the internet won outright, it took a generation. In 2005, physical formats still accounted for more than 90% of US recorded music revenue. By 2025, digital formats accounted for about 85%. The internet was widely accessible long before businesses finished changing how they operated.
Now compare that with AI.
ChatGPT reached an estimated 100 million monthly active users within just two months of its release.
Business adoption is also moving fast. According to the US Census Bureau, approximately 4% of American businesses used AI to produce goods or services in late 2023. By May 2025, that figure was approaching 10%. In less than two years, business adoption has more than doubled.
AI has several speed advantages. New capabilities can reach millions through existing infrastructure, and improvements to a single model can spread across thousands of applications, agents, and business processes. AI can also help build and integrate the systems needed to deploy those improvements.
Testing, governance, and organizational change still take time. But rapid distribution, broad reuse, and AI-assisted implementation could compress years of adoption into months.
Time-to-impact estimates vary widely. At the conservative end, MIT economist Daron Acemoglu expects AI to profitably take over only about 5% of tasks within a decade. Goldman Sachs, closer to the middle, expects about 9% of US workers to be displaced over the next ten years. At the aggressive end, Anthropic CEO Dario Amodei warned in 2025 that AI could eliminate half of entry-level white-collar jobs within one to five years.
I sit closer to the aggressive end. I believe that by the early to middle 2030s there will be fundamental changes in how most industries operate.
Scope
The second distinction is the range of work that will be affected.
Earlier technologies generally automated particular types of activity. Agricultural machinery replaced physical labor. Computers automated calculations and information processing. Industrial robots took over repetitive manufacturing tasks.
Each time, displaced workers had somewhere to go. Farmworkers moved into manufacturing and services. Manufacturing workers moved into logistics, retail, and healthcare. Clerical workers moved into new kinds of information work. There was time, and there were plenty of alternative roles.
AI complicates that process because entire clusters of related occupations could be affected simultaneously.
Consider customer-facing work: call center agents, telemarketers, travel agents, bank tellers, and receptionists.
Or administrative and information-processing roles: data entry, bookkeeping, scheduling, transcription, translation, and routine document preparation.
Software development is another example. AI can already perform entire portions of programming, testing, debugging, and user-interface development. Experienced developers will remain essential, but I expect companies to need fewer junior employees.
Exposure varies considerably, but if several related occupations are being automated at the same time, moving from one to another may no longer be viable.
Now add automation and robotics.
The International Federation of Robotics reports that more than five million industrial robots were operating worldwide in 2025. Most do defined tasks in controlled environments. Operating in unfamiliar ones is much harder, but progress is accelerating: better AI makes better robots, and better robots generate the data that improves the next generation. As costs fall, manufacturing, warehousing, transportation, food production, and facilities maintenance all become practical to automate.
Consider autonomous vehicles. Taxi and rideshare drivers are the obvious casualties. But a single capability, driving without a human, could change the economics of a whole network of industries: trucking, local delivery, public transportation, parking, vehicle maintenance, insurance, auto financing, and dealerships.
If knowledge work and physical labor are being reshaped within the same economic generation, where do displaced workers go?
Reskilling millions of people only works if governments and corporations undertake a massive effort, and I’ve yet to see any sign of one.
Size
The third distinction is the potential scale of labor displacement.
Forrester forecasts that AI and automation could eliminate 10.4 million US jobs by 2030. For perspective, the US lost 8.7 million jobs during the Great Recession. Forrester notes an important difference: recession losses are largely cyclical, while the losses it projects from automation are structural.
Goldman Sachs estimates that approximately 15 million American workers could be displaced and need to find new employment over the next decade. OpenAI's 2026 AI Jobs Transition Framework classifies 42% of US employment as facing higher automation risk or significant reorganization.
These estimates measure different things and aren't additive. They also anticipate new jobs being created as others disappear.
The question is whether those new jobs will emerge quickly enough, and in sufficient numbers, to absorb the displaced workers.
The optimists have a serious case. Goldman Sachs economist Joseph Briggs points to research suggesting that approximately 85% of US employment growth over the last 80 years came from the creation of new kinds of work. He expects that pattern to continue.
MIT economist Daron Acemoglu is less certain. As he points out, there's no law of economics guaranteeing that job creation will keep pace with job destruction.
I share that concern, and it's where I disagree with another MIT economist, David Autor.
In his 2024 essay AI Could Actually Help Rebuild the Middle Class, Autor argues that AI could make valuable expertise accessible to less-specialized workers, allowing them to perform tasks previously reserved for highly trained professionals. Research supports the productivity benefits, and Autor believes we can direct those gains toward creating valuable new human work.
But consider what happens when AI makes the average employee twice as productive and able to work at a higher level of skill and complexity.
Initially, that's an enormous advantage. Companies can grow, improve service, develop new products, or increase margins. Some will innovate and expand employment. Many will take the productivity gains, cut headcount, and celebrate the success of their AI strategy.
Now give every competitor the same capability.
For companies that don't innovate, yesterday's exceptional productivity will become the new competitive baseline. The scarcity value of expertise will decline, and companies will need fewer employees to produce the same output.
Millions of new jobs may emerge, but I don't believe we can count on them appearing as quickly as AI and automation eliminate the need for existing ones.
Big Data, Computing Power, and Attention
My fourth distinction is technical.
By the mid-2010s, decades of digitization had produced enormous amounts of data, and GPUs had made it practical to train models on it. Then, in 2017, a team at Google published Attention Is All You Need, introducing the transformer. When I read it, the implications were clear: data and computing power finally had an architecture that could put them to work at scale. That was the moment everything changed. And then it changed again.
AI is helping improve AI.
Anthropic reports that its AI systems wrote more than 80% of the code merged into its own codebase as of May 2026. In September 2026, researchers including Geoffrey Hinton and Yoshua Bengio published a paper examining how automating AI research could compress years of technological progress into months or less.
Take that a step further. What happens when AI can develop more capable successors on its own, which then develop even better ones?
That’s recursive self-improvement, or RSI. In its most advanced form, it could set off an intelligence explosion, with each generation accelerating the next.
Consider the numbers. Researchers at METR found that the length of software tasks frontier AI agents could successfully complete was doubling approximately every seven months. More recent measurements put it closer to every four.
Now imagine a self-improving system cutting that doubling time to seven days. Twenty-four doublings would mean a 16-million-fold increase in less than six months.
That’s a mathematical illustration, and intelligence doesn’t scale that neatly. Computing power, energy, and physical constraints could slow or stop it. But AI is already accelerating its own development, and the loop is tightening.
This is the road to artificial general intelligence (AGI), AI that matches people across a wide range of intellectual work, and eventually artificial superintelligence (ASI), which would exceed us. An intelligence explosion could outrun our ability to validate scientific discoveries, overwhelm human oversight, and concentrate vast economic and geopolitical power in very few hands.
The disruption starts well before AGI. AI that can reliably do a substantial share of economically valuable work, at a fraction of the cost, is enough to change how we earn, how we spend, and where we find meaning.
Where Are All the Missing Jobs?
Given everything I’ve described, why haven’t we seen mass unemployment from AI?
My 2024 paper included scenarios involving millions of displaced workers, and I expected more visible disruption by now, but I underestimated how long it would take enterprises to absorb the technology.
Legacy systems, fragmented data, regulation, and organizational resistance all slow adoption. Automating tasks doesn’t necessarily eliminate jobs, either. Companies still have to redesign how work gets done, and aging worker demographics and new demand absorb some of the impact. Even so, companies are getting faster at putting AI into production.
As recently as October, Yale’s Budget Lab found little evidence of economy-wide AI-driven disruption. But in late October, Microsoft offered voluntary buyouts to about 27% of its workforce. And Stanford researchers identified a 19% relative employment shortfall among workers aged 22 to 25 in highly AI-exposed occupations, mostly through reduced hiring.
Entry-level jobs are how people become experienced professionals. If AI takes over that work, where do the senior professionals of 2036 come from?
We may be removing the first rungs of the career ladder while continuing to depend on the people who climbed it.
To quote The Replacements, “God, what a mess, on the ladder of success, where you take one step and miss the whole first rung.” The song is about alienation and a new generation struggling to fit in. At least the first rung was there. Now it’s gone.
The Darker Economic Scenario
I wouldn't dismiss US unemployment reaching 10% to 15% during a severe AI-and-automation transition. That's a stress scenario, with considerable uncertainty. For perspective, unemployment is currently 4.2%. It reached 10% following the Great Recession and 14.8% in April 2020. Both shocks were followed by employment recovery. What concerns me is a structural decline in demand for human labor that may not reverse.
Years ago, while researching how to hedge mortgage portfolio risk at Bank of America, I learned something that stuck with me. Consumer spending drives roughly two-thirds of the American economy.
Today, it's about 68%. A 10% decline in consumer spending would represent a demand shock equivalent to roughly 6.8% of GDP, before secondary effects or offsets. Government spending and business investment could compensate for some of that loss, but there's no obvious replacement for consumer demand at that scale.
And the effects could compound. Job losses reduce household income. Consumers spend less. Businesses lose revenue, cut costs, and eliminate more jobs.
The consequences reach banks, insurers, and governments. Credit losses rise, insurance markets tied to employment and payroll change, and tax revenues could decline just as demand for assistance increases.
Banks that anticipate these changes can adjust underwriting, reserves, and products. Those that don’t could face heavy losses and pull back on lending just when people and businesses need it most.
There's another risk: where the wealth goes. If AI allows companies to produce more with fewer employees, a larger share of the economic gains could flow to owners and investors rather than workers. That wealth is already concentrated. The wealthiest 10% of Americans own about 88% of corporate equities and mutual fund shares, while the bottom half owns less than 1%. As AI shifts more economic value from labor to capital, we could see extraordinary wealth creation increasingly concentrated in relatively few hands.
We could end up with an economy capable of producing more wealth than ever, while millions of people lose their livelihoods, their financial security, and a fundamental part of how they define themselves.
And Then What?
One of the essays that most influenced my thinking was My Last Five Years of Work, written in 2024 by Avital Balwit, then 25 and working at Anthropic.
She wondered whether AI might soon make her job unnecessary, and whether people could live fulfilling lives without work.
Looking at research on unemployment, retirement, and well-being, she found an important distinction. Losing a job can be devastating. Having the resources and freedom to choose how to spend your time can be liberating.
Balwit imagined a highly automated future in which people’s material needs were met through universal basic income or other transfers.
We haven’t built that system. Still, imagine what’s possible. Real abundance, affordable goods and services, major scientific advances, and far greater freedom over how we spend our lives.
I’d happily spend more time playing music, swimming, cycling, and doing the other things that make life interesting.
But there’s another problem that money alone can’t solve.
In America, we often ask “What do you do?” when we’re really trying to understand who someone is.
Our careers become part of our identities. Francis Fukuyama argued that the desire for recognition, achievement, and the opportunity to distinguish ourselves is a fundamental human drive. We want to contribute, to be good at something, and to be respected for what we’ve accomplished. We want to feel needed.
For many Americans, work provides that recognition.
Some will find it elsewhere, in family, art, sport, or service. Others may struggle, particularly when losing their work wasn’t their choice.
AI and robotics may eventually allow us to produce much of what we need with considerably less human labor. I believe that possibility is real, and that it could arrive far sooner than most people expect.
The technological opportunity is extraordinary. The economic and social transition could be messy.
We’ve spent centuries trying to build machines that can do our work.
What happens if we finally succeed?