AI Could Push 11 Million U.S. Workers Into New Careers by 2035, Study Finds

AI Could Push 11 Million U.S. Workers Into New Careers by 2035, Study Finds

TL;DR

Institutional projections now point in the same direction: roughly 11 to 12 million U.S. workers will need to change occupations because of AI by the early 2030s. The productivity gains are real and documented. What none of the major studies answers honestly is whether the new careers pay as well as the ones they replace.

Key Takeaways

• McKinsey Global Institute's July 2023 report found that approximately 11.8 million U.S. workers may need to change occupations by 2030, with generative AI compressing timelines that pre-pandemic projections placed a decade further out.
• Goldman Sachs estimated in 2023 that generative AI could expose the equivalent of 300 million full-time global jobs to automation, with roughly 18 percent of work potentially automatable entirely — a figure the bank paired with a 7 percent global GDP upside that received considerably more coverage.
• The World Economic Forum's Future of Jobs Report 2023 projected 83 million jobs eliminated and 69 million created over five years — a net contraction of 14 million, almost never cited in the headlines that covered it.
• OECD analysis has consistently documented that workers in high-automation-risk roles earn significantly less than those in low-risk roles — meaning the workers most exposed to displacement are also those least financially equipped to fund a transition.
• MIT economist Daron Acemoglu has argued in multiple papers that automation creates fewer new tasks than it destroys in the near term, and that AI-driven productivity gains concentrate in tasks and firms, not across the broader workforce.
• Bureau of Labor Statistics projections through 2033 show employment growth concentrated in healthcare support, personal services, and skilled trades — roles requiring physical presence, not the knowledge work that AI targets most directly, and roles that often pay less than the displaced positions they absorb.

The Number Is Not the Story

Eleven million sounds precise. It isn't.

It sits inside a range that every major institutional labor forecast published in the last two years shares. McKinsey puts the U.S. figure at 11.8 million by 2030. The World Economic Forum's math implies a net contraction of 14 million. Goldman Sachs doesn't count jobs at all — it counts tasks exposed, which is a softer metric, and one that happens to be more palatable to the clients who buy the research.

What the numbers share is a direction. The labor market is heading toward significant occupational churn, faster than prior waves of automation because generative AI targets cognitive work, not just manual work. Paralegals. Marketing coordinators. Junior financial analysts. Medical coders. The roles that absorbed college graduates from the 1990s through the 2010s are the ones under the most immediate pressure.

So the 11 million figure is a reasonable central estimate. The word that should make journalists and policymakers pause is "careers." Who defines what counts as a new career? The model does. Transition from a $52,000-a-year paralegal role into a $38,000 healthcare support job is, in the data, a successful transition. In lived experience, it is a 27 percent wage cut.

What the Data Actually Shows

The Sectors With the Largest Exposure

McKinsey's analysis identified office and administrative support as the highest-exposure category — millions of U.S. workers in roles where more than 70 percent of tasks are potentially automatable by the decade's end. Customer service, data processing, and administrative coordination lead the list.

Behind those: business and legal support roles, and transportation and logistics, where route optimization and scheduling automation are already moving through dispatch and coordination work.

The sectors least exposed are healthcare delivery, construction, and personal services. Not because AI hasn't arrived — it has — but because the work requires physical presence or real-time human judgment that current systems cannot reliably replicate at scale. These are also the sectors the BLS projects will absorb displaced workers. The gap between what those roles pay and what the displaced roles paid is not a rounding error.

The Wage Gap No Projection Quantifies Cleanly

This is the structural gap in the research. Every major forecast counts workers who will need to transition. Almost none track what those workers earn on the other side.

OECD data shows, consistently, that high-automation-risk occupations cluster at lower wage bands. The workers most exposed to AI displacement are concentrated in the bottom three income quintiles. Their access to funded retraining programs, geographic mobility, and financial cushion during transition gaps is also lowest. That is not a coincidence — it is the same economic stratification that made those roles automatable in the first place. Employers automated the tasks they could because the labor was replaceable. The workers in those roles were replaceable because they lacked the bargaining power to make themselves otherwise.

Daron Acemoglu's framework is useful here. His argument — developed across several MIT and NBER papers — is that the new tasks created by automation are predominantly high-skill tasks. They require different education profiles, different credentials, different social capital. The worker displaced from a claims-processing role does not automatically have a path into the prompt-engineering or AI-oversight role that partially replaced it. That mismatch is not a transition problem the market solves cleanly.

Goldman Sachs's 7 percent GDP projection doesn't contradict any of this. GDP can rise while median wages stagnate. That happened throughout most of the 2010s, in precisely the economy where AI investment was first accelerating.

Institutional Projections Compared

The methodology gap matters. Task-exposure models count what is technically automatable. Employment transition models count what actually happens in labor markets. The two are not the same, and the distance between them is where workers fall through.

What This Changes for Journalists, Policymakers, and Labor Economists

For Journalists

The newsroom is not a bystander here.

AI is already performing tasks that, five years ago, were junior reporter assignments: earnings summaries, sports scores, local government meeting recaps, product description copy. The Reuters Institute for the Study of Journalism's Digital News Report 2024 documented growing AI adoption across news organizations — against a backdrop of continued editorial headcount reductions. The correlation is not proof of causation. The timing is not coincidental.

The accountability gap is worth naming: the same publications that front-page "AI will reshape the labor market" are often the organizations deploying AI to reduce their own labor costs, beginning with the entry-level roles that once trained the mid-level roles. Venture capital continues to concentrate in AI tools that automate content creation and knowledge work — including roles that have historically employed the mid-career professionals those same publications hire.

Covering AI's labor market effects without covering your own industry's adoption of AI is a significant blind spot. The story is not only about manufacturing plants and call centers. It is about the desk adjacent to yours.

For Policymakers

The EU AI Act, which entered into force in August 2024, mandates transparency for high-risk AI systems — including those used in employment decisions. It does not mandate retraining funds, wage replacement during transitions, or minimum outcome standards for workers displaced by covered systems.

That gap is the current policy problem. The U.S. has no equivalent framework at the federal level. Workforce development programs — federal and state — were designed for manufacturing-era transitions. They assume workers need to acquire new physical skills, not compete with systems that already outperform them at producing documents, summarizing legal text, classifying insurance claims, and generating first-draft marketing copy.

The political economy runs in one direction: the firms that deploy AI capture the productivity gains immediately; the workers who absorb the displacement costs are distributed across geographies, sectors, and decades. Concentrated benefits, diffuse costs — the same structural problem as every previous round of capital-labor conflict over automation. The policy response has historically been slow, underfunded, and arrived too late to help the first wave of workers affected.

For Labor Economists

The methodological constraint is measurement lag. Task-automation models tell us what is exposed; employment surveys tell us what happened — usually 18 to 24 months later, at best. By the time BLS data reflects AI-driven displacement at scale, the technology will have moved another generation and the policy window will have narrowed.

The more urgent research question is not "how many jobs" but "which workers," "at what wages," and "who funds the transition gap." The OECD has begun producing transition-pathway analyses. MIT's Work of the Future task force has called for longitudinal tracking of displaced workers by name, not just by occupational category. Neither effort has yet produced the dataset that policymakers actually need to design effective intervention.

How to Read the Next AI Labor Projection

Before citing a forecast in a brief, an article, or a hearing:

• Does it count tasks or jobs? Task exposure overstates displacement — most jobs contain tasks that can't be automated alongside tasks that can. The two numbers are not interchangeable.
• Does it address transition wages? A study that says workers will "move into new careers" without tracking wage outcomes is describing movement, not outcomes.
• Who funded it? Projections from consulting firms with active AI advisory practices, AI platform vendors, or think tanks funded by major technology companies carry a directional incentive worth disclosing.
• What's the time horizon? "By 2030" and "by 2035" produce meaningfully different retraining cost estimates and wage-erosion curves. Treat them as different claims.
• Does it disaggregate by income? Aggregate projections that don't break down impact by wage decile conceal the fact that low-wage workers face higher automation risk and lower reskilling capacity simultaneously.
• Does it document actual transition outcomes? Modeling what will happen is not the same as documenting what has already happened in earlier automation cycles. Projections built solely on task-automation rates are forecasts, not evidence.

Where This Is Heading

Reskilling will remain structurally underfunded. The U.S. spends approximately 0.1 percent of GDP on active labor market programs — compared to 0.5 to 1 percent across Germany and the Nordic economies. That gap predates AI. It will widen as the volume of workers needing retraining increases.

The EU AI Act will generate labor impact data the U.S. currently lacks. High-risk AI system audits under the Act will require documentation of employment effects across covered systems. The disclosure framework is imperfect and slow-moving — but it will produce better evidence than anything currently available in the U.S. market.

Sector-level bargaining will expand. The Screen Writers Guild negotiation over AI in Hollywood was a preview. Similar disputes are active in journalism, financial services, and healthcare. Those negotiations will produce precedents — on AI disclosure requirements, on residual payments, on retraining obligations — that matter more than any projection.

The productivity gains are real and the distribution of those gains is the policy question. Goldman Sachs's 7 percent GDP projection is probably directionally correct over a ten-year horizon. Post-automation productivity gains from the 1980s manufacturing transition took 15 to 20 years to show up in median wages — and then only partially. There is no structural reason to expect this transition to distribute differently.

The measurement infrastructure needs rebuilding. Quarterly employment surveys were not designed to track occupational transitions at AI's deployment speed. The Fed's regional banks and several university-based labor economics programs are building better instruments. They are running at least two years behind the technology.

FAQ

The headline says 11 million. Other studies say 12 million, 14 million, or 300 million. Why do the numbers vary so much?

They're measuring different things. The 11 to 12 million figure counts U.S. workers likely to need full occupational changes. The 300 million figure counts global jobs with a significant share of tasks exposed to automation — a much lower bar than requiring a career switch. The 14 million is a net global figure, not a U.S. figure. The methodologies are not comparable and should not be added together.

If AI creates new jobs, isn't this net positive for the labor market?

Possibly, in aggregate and over time. The WEF's math — 83 million eliminated, 69 million created — is a net loss. McKinsey is more optimistic about job creation. Neither study has tracked whether the workers displaced are the same workers who fill the new roles. Historical evidence from prior automation waves strongly suggests they are not, particularly when new roles require different credentials or different geographic locations.

Isn't this just the latest version of a fear that never materializes? Workers adapted before.

Partly right, and partly where the historical analogy breaks down. Previous automation waves primarily displaced routine manual work and created more non-routine cognitive work. This wave is targeting non-routine cognitive work directly. The workers who historically absorbed prior waves of displacement by moving into knowledge roles are now in the category most directly affected.

What should government actually do?

In the immediate term: longitudinal tracking of displaced workers — not just employment counts, but wage outcomes over three to five years. Wage replacement during retraining periods, not access to courses workers must fund themselves. The EU AI Act creates a transparency framework; it does not create a safety net. The U.S. needs both.

How should individual workers in high-exposure roles interpret this?

Skeptically but not dismissively. "Technically automatable" and "actually automated in your employer's budget this year" are not the same thing, and the timeline between them is longer than headlines suggest. But the direction is clear enough that waiting for certainty before assessing skill adjacencies is a strategic mistake.

What's the single most important thing missing from every projection published so far?

Wage outcomes at the individual level. Every major forecast counts transitions; almost none track whether the new role pays comparably to the one it replaced. That is not a minor omission — it is the number that would determine whether "AI creates new careers" is a reassurance or a euphemism.

Who is best positioned to produce the research that's actually needed?

The Federal Reserve's regional banks have begun producing district-level survey data on wage and transition outcomes that national models miss. BLS, in partnership with OECD and academic labor economists, needs a longitudinal survey infrastructure that tracks individual workers across transitions — not just counts occupational categories before and after. That work requires sustained public funding and a willingness to publish findings that don't support optimistic policy narratives. Neither condition is currently guaranteed.