Why no one can agree on what AI will do to jobs
Will AI repeat history — or break it?
People building AI think it will eliminate many millions of jobs. People who study labor markets think it won’t. At least one of these groups is badly wrong — and the stakes are extremely high.
Dario Amodei, CEO of Anthropic, made headlines last spring when he predicted that AI could wipe out half of all entry-level white-collar jobs — and spike unemployment to 10-20% in the next one to five years. “Most of them are unaware that this is about to happen,” he said. “It sounds crazy, and people just don’t believe it.”
The warnings don’t just come from tech CEOs. Bernie Sanders and Geoffrey Hinton have also joined the chorus warning of impending mass unemployment. And voices that might be expected to be far more restrained, such as Federal Reserve Chair Jerome Powell, have also acknowledged the possibility. In past technological transitions, Powell noted, jobs were destroyed and others created, ultimately yielding higher productivity and enough jobs for people. “This may be different,” he said last month.
However, these forecasts stand in stark contrast to the steady predictions coming from established sources in banking and economics, such as Goldman Sachs, McKinsey, the IMF and the World Economic Forum. Their models foresee rising AI-assisted productivity, with a small unemployment increase during the transition period — not surging structural unemployment. A Goldman Sachs report from August succinctly captures the calm assurance: “We remain skeptical that AI will lead to large employment reductions over the next decade.”
And the gap between the two camps is enormous. Amodei predicts 10-20% unemployment. Goldman Sachs expects something closer to a half–percentage-point increase while workers adjust.
For context, the US unemployment rate was under 5% last year, and the Great Depression is the only time in recorded history that the unemployment rate lingered above 10% for an extended period. A sustained unemployment rate in the teens would represent an economic rupture on a scale not seen in recorded history.
And which of the forecasts best resembles the future has huge implications for other questions we need to ask ourselves. At 20%, how society copes with the many people who find themselves economically disenfranchised — as recently discussed by Dwarkesh Patel and Phil Trammell — suddenly becomes a burning issue. A 0.5% increase is a manageable problem.
Beneath these forecasts lie two distinct disagreements: how capable AI will become, and how quickly it can or will diffuse into the economy.
Are we already seeing the first cracks?
There are early signs disruption has already started.
A Stanford study published in November found a 13% decline in early-career jobs most exposed to AI. Reports of AI-driven automation are appearing with increasing frequency in the Federal Reserve’s Beige Book, which tracks anecdotal evidence about economic conditions. For example: “Many contacts noted that even modest deployments of AI would enable them to not refill some jobs or to skip a recruiting class of entry-level workers.”
And more and more anecdotal accounts are popping up. Anton Leicht, a visiting scholar with the Carnegie Endowment’s Technology & International Affairs team, has spoken to people in markets such as the Philippines, where AI-powered augmentation has allowed firms to lower prices, increasing competitive pressure on US companies. More informally, stories of job losses attributed directly to AI are becoming commonplace — including, in my own small circle, a relative who lost her job because “AI could do it worse for free.”
Public sentiment reflects this unease. Pew surveys show that a majority of American adults expect AI to reduce the number of available jobs. A survey of hundreds of industry experts finds that 75% anticipate slowing white-collar job growth, or outright losses, by 2030. Even the most optimistic forecasts assume widespread reskilling and significant changes to how work is done.
At the same time, unemployment has not spiked. So far, the data is consistent with a normal technological transition — or the early stages of something far more disruptive. The question is which trajectory we’re actually on.
The case for “technology as usual”
Repeatedly through history, new automation technologies — the printing press, industrial revolution, steam engine, electricity, the internet, ATMs, agricultural mechanization — were hailed as destroyers of jobs that would cause mass unemployment.
They didn’t.
New technologies made some jobs obsolete and caused real hardship for displaced workers. But over time, productivity gains created new industries, new roles, and higher average living standards. For people entering the workforce, the new economy often offered more opportunities than the old one.
Moderate forecasts for AI lean heavily on this track record. A McKinsey Global Institute report argues that “technological advances often cause disruption, but historically, they eventually fuel economic and employment growth.” An 2023 OECD report found that “the consensus among economists and policy makers from previous rounds of automation technologies…is that labour demand should remain strong,” but the scale of challenge may be bigger this time, requiring new policies to help navigate the transition. Goldman Sachs made the point more bluntly last summer: “Predictions that technology will reduce the need for human labor have a long history but a poor track record.”
Notably, there’s actually little disagreement about the above moderate scenario happening in the short term. Even Amodei’s dramatic predictions imply an initial phase of augmentation and partial automation: fewer entry-level hires, some redundancies, and growing pressure to reskill.
The deeper disagreement is what will happen after that initial disruption phase: will AI settle into a steady augmentation state, as past technologies did, or will it continue advancing until it can perform essentially all economically valuable work?
Why AI progress might slow (enough)
Many economists remain fundamentally skeptical that AI will threaten all jobs anytime soon — or ever. Their skepticism tends to rest on several potential bottlenecks.
It is possible that AI will never be able to perform all human tasks. Large parts of the economy remain difficult to automate, particularly roles that require robust robotics, operation in unpredictable physical environments, or multi-step agency. As Kweilin Ellingrud, a senior partner and director at the McKinsey Global Institute, puts it: “There are huge swaths of the economy where I don’t think generative AI is going to even majority — much less completely — replace the work we need for a long time, if ever.”
There are also natural speed limits to adoption. Sensitive tasks require trust and reliability, which take time to establish. Organizations need years, not months, to integrate new technologies into workflows. Real-world data collection cannot be sped up arbitrarily. For example, electricity provided a huge technological disruption that touches almost every part of the economy today — but it took decades to realize that impact. Businesses had to redesign warehouses, processes, and workflows before they could fully harness the new productivity gains.
“As promising as AI may be, there’s little chance it will live up to [the] hype,” MIT economics professor Daron Acemoglu said in late 2024, estimating that AI would only be ready to take over, or even heavily aid, in around 5% of jobs over the next decade.
Under this view, AI may eventually automate much of the economy — but only over many decades. That slower timeline would allow labor markets and institutions time to adapt, muting the worst disruptions.
What if AI doesn’t stop improving?
Other economists argue that models treating AI as a one-time shock — like electricity or the steam engine — are making a strong and potentially mistaken assumption. Unlike past technologies, AI systems have shown sustained, rapid improvements across a wide range of cognitive tasks, points out Tom Cunningham, an economist at AI evaluations non-profit METR, “and we don’t know where the ceiling is.”
Leicht captures the concern succinctly: if AI keeps improving quickly enough, it may end up “chasing everyone who’s chasing jobs,” acquiring new skills faster than displaced workers can retrain.
Even if we invent new jobs, economics professor Anton Korinek tells me, intelligent machines will be able to learn and perform those new jobs more efficiently than human workers. That would leave “nothing to switch to.” Even if humans retain a comparative advantage in some niches, wages could be driven down below livable levels if AI labor is dramatically cheaper — possibly leading to surging inequality.
Peter Wildeford, chief strategy officer at the Institute for AI Policy and Strategy, compares the potential shift to the industrial revolution’s effect on physical strength. Once machines became stronger than humans, brute strength ceased to be an economically valuable trait. AI, he argues, could do the same for intelligence.
What policy can do under deep uncertainty
If hard bottlenecks prevent AI from surpassing human capabilities in key domains, the historical analogy may hold. We could see a period of adjustment followed by a new, stable equilibrium of human–AI collaboration.
But if unemployment begins rising sharply, it may signal that more profound disruptions are coming. In that case, policy responses will matter enormously. Proposals range from expanded unemployment insurance and large-scale retraining to universal basic income or broader redistribution of AI-generated wealth.
In the short term, this creates a difficult balancing act. AI-driven change is likely to be rocky, leaving many people worse off — at least temporarily — and that harm needs to be addressed.
But policies that attempt to shield workers by restricting AI augmentation risk sacrificing many benefits. In the world where AI remains a useful tool, it could contribute to substantial beneficial progress. As Kweilin Ellingrud and Anu Madgavkar of the McKinsey Global Institute tell me, AI-assisted productivity growth may be the only available path to sustaining economic growth and improving living standards in the face of a shrinking population.
And if AI ultimately becomes capable of fully automating work, one country limiting its deployment could leave human workers especially vulnerable, outcompeted by automated systems elsewhere in the world.
For now, the safest course may be to focus on measures that address present disruptions without foreclosing future options if mass displacement later becomes reality: strengthening social safety nets, supporting retraining, encouraging AI fluency in schools, and monitoring labor market indicators closely.
Most likely, we have several years of rocky transition before we might confront the possibility of mass automation — and we may see signs during that period that AI progress is slowing. If investment in AI infrastructure drops off dramatically, for example, that may be an early sign that fears of total job loss were overstated.
Perhaps, in a few decades, we will look back on today’s warnings as yet another case of technological alarmism. Or perhaps we will see them as unusually prescient. Wildeford expects it will be the latter, as he warns that it “seems really difficult to avoid unemploying the entire American workforce.”
For now, all we can do is watch carefully — and prepare for the possibility that this time really is different.
 | A guest post by
|
One thing I really appreciate about Transformer is how you get to the core of a problem in AI and give equal weight to all points of view on it. This article is a great example; another is the article you guys did about how generative AI is challenging the original intention of copyright law. You guys do great reporting.
The best way to survive in 2026:
It’s to do what AI can’t. Ask a novel question.