What the Data Actually Says — and Why Neither Side Is Telling the Full Story
Every few months, the cycle repeats. A major tech company announces layoffs and credits AI. A breathless headline follows: “AI Will Destroy Millions of Jobs.” Then a think tank publishes a counter-report: “AI Creates More Jobs Than It Kills.” Social media splits into two camps, neither citing a single primary source.
This piece is not going to do that.
Instead, we’re going to do something simple but uncomfortable: read the actual data. Government labour projections, IMF assessments, Goldman Sachs research, BLS occupational statistics — the kind of sources that testify before Congress and move bond markets. When you read them carefully, a more nuanced picture emerges. One that frustrates the doomsayers and the optimists in roughly equal measure.
The Claim Everyone Is Arguing About
The popular framing goes like this: AI is either going to destroy the economy or supercharge it. Dario Amodei, CEO of Anthropic, warned in early 2025 that AI could “eliminate essentially all” entry-level white-collar jobs within a few years. Jensen Huang of Nvidia fired back at VivaTech 2025, arguing that greater productivity historically leads to more hiring, not less. Two of the most powerful people in AI can’t agree on which direction the sword is pointed.
The public debate, stripped of its emotion, is really about two different questions that people keep conflating:
- Will AI destroy specific jobs? (Almost certainly yes, in some sectors)
- Will AI destroy employment overall? (The data says: probably not — but the transition will be deeply unequal)
These are not the same question. Confusing them is how we end up with bad policy, bad career decisions, and bad discourse.
What the Institutional Data Actually Shows
Let’s start with the largest, most credible datasets available.
The World Economic Forum: A Net Positive, But a Massive Disruption
The WEF’s Future of Jobs Report 2025 — drawn from surveys of over 1,000 employers representing more than 14 million workers across 55 economies — produced the most widely cited projection in the current debate. By 2030, the report projects 92 million roles will be displaced while 170 million new roles emerge, resulting in a net gain of 78 million jobs.
Source: World Economic Forum, Future of Jobs Report 2025 — Survey of 1,000+ employers, 14M+ workers across 55 economies
That headline number gets misused constantly. The “net gain of 78 million” gets treated as proof there’s nothing to worry about. But read the fine print: 41% of employers globally plan to reduce their workforce in areas where AI can automate tasks. That’s not a distant future number. That’s what senior decision-makers said they were planning right now. The 78 million net gain is an aggregate; the individual pain of displacement is not distributed evenly, and the people who lose routine jobs are rarely the same people who get hired for AI-adjacent ones.
Source: Goldman Sachs (2023) — % of tasks in each sector automatable by current generative AI
The IMF: 40% of Global Jobs, 60% in Advanced Economies
In January 2024, the IMF released one of the most rigorous analyses of AI’s labour market exposure to date. Their finding: roughly 40% of jobs globally face meaningful exposure to AI capabilities. In advanced, digitised economies — the United States, Europe, Japan — that number climbs to approximately 60%.
The word “exposure” is doing a lot of work here. Exposure doesn’t mean elimination. The IMF explicitly distinguishes between jobs where AI will substitute for human labour and jobs where it will complement it. In the complementary scenario, AI makes the worker more productive — which could mean higher wages, but also fewer total workers needed to produce the same output.
What the IMF is most concerned about is inequality. Their analysis found that AI could widen the gap between high-income workers (who can leverage AI to become more productive) and low-income workers (who are more likely to be displaced without the skills to pivot). IMF Chief Kristalina Georgieva’s blog post accompanying the report was unusually direct: “The AI era is upon us. It is still within our power to ensure it brings prosperity for all.”
Goldman Sachs: 300 Million Exposed, But “No Discernible Impact” Yet
Goldman Sachs entered this debate in March 2023 with a report that generated headlines across every major publication: up to 300 million jobs globally are exposed to automation by AI. In the U.S. specifically, two-thirds of occupations could see at least some share of their tasks automated, with office and administrative support jobs leading at 46% task automation potential, followed by legal (44%) and architecture and engineering (37%).
But here’s the part that rarely makes the headlines: Goldman’s own economists, reviewing the data in early 2025, concluded that despite rapid AI proliferation, AI has had “no discernible impact” on major labour market indicators — unemployment rate, layoffs, or productivity measures — as of their analysis. Economist Joseph Briggs, who co-leads Goldman’s global economics team, put the timeline for broad AI labour displacement at roughly 10 years, with only 6–7% of workers displaced during that transition window in their base case.
Goldman’s own research contradicts the panic its earlier report helped generate.
The Bureau of Labour Statistics (BLS): The Numbers Governments Bet On
The BLS publishes the most technically rigorous employment projections in the world — the numbers that inform government budgets, education policy, and workforce training programmes. Their 2023–2033 projections, updated in 2024, incorporate AI as a structural factor.
The findings are specific and worth reading carefully:
- Software developers: projected to grow 17.9% from 2023 to 2033 — “much faster than average”
- Database architects: projected to grow 10.8%
- Data scientists, actuaries, operations research analysts: all projected to grow at least 20% by 2034
- Nurse practitioners: projected to grow 52%
- Bank tellers: projected to decline 15% (approximately 51,400 jobs eliminated)
- Cashiers: projected to decline 11% (353,100 jobs)
- Customer service representatives: projected to decline 5%
- Graphic designers: employment growth expected to be “limited” due to generative AI
The BLS picture is not “AI destroys all jobs.” It’s a story of bifurcation: jobs requiring complex human judgment, technical depth, and interpersonal skill are growing. Jobs consisting primarily of routine, rule-based, or easily replicable tasks are shrinking.
That distinction — routine vs. non-routine — is the most important thing in this entire debate, and it barely appears in mainstream coverage.
The Routine Reality: It’s Not About AI. It’s About Tasks.
Economists have known for decades that automation doesn’t eliminate jobs uniformly. It eliminates tasks. The question is whether the remaining tasks — the ones automation can’t replicate — are valuable enough to sustain employment at the same or better levels.
The academic framework underpinning all serious research here is the task-based model developed by economists including MIT’s Daron Acemoglu (co-winner of the 2024 Nobel Prize in Economics). In this framework, automation creates three effects simultaneously:
- A displacement effect: it substitutes for human labour in specific tasks
- A productivity effect: it makes human workers more efficient in remaining tasks, increasing demand for their output
- A reinstatement effect: it creates entirely new tasks — and new occupations — that didn’t exist before
The net outcome for employment depends on which of these three forces dominates. And historically, the reinstatement effect has been powerful enough to offset most displacement.
The ATM Paradox: The Most Important Case Study You’ve Been Misreading
When ATMs were rolled out across the U.S. starting in the 1970s and accelerating through the 1990s, the assumption — inside banks and outside them — was that bank teller jobs would disappear. It seemed inevitable. Machines now dispensed cash and took deposits. What was left for the human?
What actually happened is one of the most cited examples in labour economics, and it was extensively documented by economist James Bessen in IMF Finance & Development. The number of ATMs in the U.S. grew to over 400,000 by the mid-2000s. And yet: the number of bank teller jobs did not fall. They grew.
Source: IMF Finance & Development (Bessen, 2015); BLS Occupational Employment Statistics — Illustrates the “automation paradox”
Here’s why. ATMs reduced the cost of operating a bank branch significantly — the average urban branch went from needing about 21 tellers to 13. This made opening new branches economically viable in smaller towns and underserved areas where banks had never operated. As branch networks expanded, total demand for tellers increased faster than automation reduced demand per branch. Meanwhile, the nature of the teller job evolved — from cash handling to relationship banking, financial advice, and complex problem-solving. Banks started hiring college graduates for teller roles.
The automation paradox: ATMs simultaneously destroyed the tasks that defined teller jobs and increased the total number of teller jobs.
There’s a critical footnote to this story that Jacob Morgan’s analysis on X in 2025 highlighted sharply: the ATM transition played out over roughly 40 years. The Jevons Paradox — where efficiency gains increase rather than decrease total consumption — had time to operate. Banks could reorganise, workers could retrain, new markets could emerge. Mobile banking, arriving fully formed around 2010, finally produced the teller decline that ATMs never did — because it was a more complete form of automation, not incremental.
This is the key difference that should terrify policymakers when thinking about AI. AI is not the ATM. It is mobile banking.
Where the Pain Is Real: The Occupations on the Front Line
Across every credible dataset, the same categories of jobs emerge as genuinely vulnerable. Not “exposed” — actually declining.
Administrative and clerical work is the most consistent casualty. Procurement clerks, credit authorizers, non-medical secretaries — the BLS explicitly projects “decline or little change” for these roles through 2033. The productivity gains from AI tools are simply reducing headcount without creating enough adjacent demand to offset.
Customer service is seeing direct displacement. Site Selection Group documented a decline of approximately 80,000 customer service positions in the U.S. between 2022 and 2024. AI chatbots and automated resolution systems have improved rapidly enough to handle a large portion of routine service interactions.
Graphic design and media production are under pressure from generative AI. The BLS noted graphic designers as an occupation “whose tasks have a high potential to be automated” and projects limited employment growth as a result.
Medical transcription (-4.7% projected through 2033) and broadcast announcing (-declining, per BLS) are both cited specifically as AI impact occupations.
The pattern across all of these is identical: they consist primarily of routine, replicable tasks that don’t require complex judgment, unpredictable physical presence, or deep social interaction. They are the cognitive equivalent of the physical tasks that assembly-line robots displaced in the 1980s.
Source: U.S. Bureau of Labor Statistics, 2023–33 Occupational Employment Projections (Aug 2024)
Source: IMF “AI Will Transform the Global Economy” (Jan 2024); ILO Generative AI & Jobs (2023)
Where the Optimism Is Justified (With Caveats)
The data is not uniformly grim. Several forces genuinely support the “AI creates jobs” side of the argument.
The AI infrastructure buildout is generating employment directly. Goldman Sachs analyst Evan Tylenda estimates that roughly 500,000 net new jobs in the U.S. alone will be needed to build and maintain the power and data center infrastructure required to sustain AI by 2030. These are electricians, lineworkers, construction workers, and engineers — not knowledge workers.
Healthcare is growing aggressively, and AI is mostly a complement there. Nurse practitioners are projected to grow 52% by 2033. Healthcare and social assistance added more jobs than any other sector in the BLS 2024–2034 projections, driven largely by demographic demand. The ILO found that in healthcare specifically, AI functions primarily as an augmentation tool rather than a replacement mechanism.
STEM employment is expanding relative to the overall workforce. According to National University analysis drawing on BLS data, the share of U.S. jobs in STEM fields grew from 6.5% in 2010 to nearly 10% in 2024 — almost a 50% relative increase.
Skilled trades are among the safest occupations in the entire economy. Installation, repair, maintenance, and construction occupations have AI automation exposure in the low single digits (4–6% per Goldman Sachs). These are jobs that require physical dexterity in uncontrolled environments, on-the-spot problem solving, and presence — none of which AI can replicate at scale with current or foreseeable technology.
The Acemoglu Counterpoint: Don’t Trust the 7% GDP Promises
Not everyone in serious economics accepts the optimistic projections. Daron Acemoglu, whose Nobel Prize-winning work on institutional economics and technological change makes him one of the most credible voices in this debate, published a paper in 2024 arguing that the productivity and GDP impacts from AI will be far more modest than the headline projections suggest.
Where Goldman Sachs projected a 9% increase in productivity and 6.1% increase in GDP over 10 years from generative AI, Acemoglu’s estimate was roughly 0.5% productivity growth and 1% GDP growth over the same period. His argument: the 5–10% of tasks that AI can genuinely automate well in the near term are not the most economically valuable tasks, and the “easy” productivity gains will be captured quickly while the hard structural transformation takes far longer.
Acemoglu told Goldman Sachs directly in an interview: the forecast differences are “primarily about timing.” He doesn’t deny AI will be transformative — he’s skeptical it will be transformative on the timeline that investors and policymakers are pricing in.
George Mason economist Tyler Cowen pushed back, arguing that Acemoglu underweights AI’s ability to accelerate scientific discovery and business innovation. Both are credible economists. Both disagree substantially on the magnitude and timing of impact.
This is not a settled debate even among the people who do this for a living.
The Inequality Problem: The Data’s Most Uncomfortable Implication
Here is what every major institutional report agrees on, and what almost nobody in the mainstream AI jobs debate talks about honestly: the pain of AI displacement is not going to be evenly distributed.
The IMF’s analysis found that older workers are at significantly higher risk of being unable to adapt to AI-related skill demands. Workers with lower educational attainment are disproportionately concentrated in the routine-task occupations most vulnerable to automation. The ILO noted that in low-income countries, only 0.4% of jobs are at risk from generative AI in its current form — compared to 5.5% in high-income countries. The same technology that is mostly harmless for subsistence agriculture is devastating for administrative services.
Within advanced economies, the BLS data and WEF projections suggest a clear stratification: workers who can ride the AI wave (tech, healthcare, finance, skilled trades) will see rising wages and growing employment. Workers in the displaced categories (admin, clerical, customer service, low-skill data processing) will face a labour market that is contracting and offering lower wages for the roles that remain.
PwC’s 2025 AI Jobs Barometer found that workers with demonstrable AI skills earn a 25% wage premium over peers without those skills. That premium is a signal — it says the market is already pricing in the scarcity of AI-adjacent competence and the surplus of AI-replaceable labour.
What History Actually Teaches Us
Every major technology wave since the Industrial Revolution has generated fears of mass unemployment that never quite materialised in aggregate — even when they caused devastating local disruptions.
Power looms automated 98% of the labour needed to weave a yard of cloth in the 19th century. The number of weavers subsequently increased for decades, because cheaper cloth created more demand for cloth. The steam engine, the dynamo, the computer — each produced a wave of forecasts that the economy was about to run out of things for people to do. Each time, new tasks appeared that nobody had predicted.
As Deloitte’s chief UK economist Ian Stewart’s research on two centuries of UK census data found: “Paradoxically, many of the fields most transformed by technology have produced the biggest increases in employment, from medicine to management consulting.” The sectors most disrupted by technology have often become the largest employers.
But history also teaches us something less comforting: the transition costs are real and they are not shared equally. The handloom weavers who lost their livelihoods in Lancashire in the 1820s did not benefit much from the long-run economic growth that followed. Many never recovered. The fact that the economy eventually finds new equilibrium is cold comfort to the specific workers caught in the disruption.
The Verdict: What the Data Concludes
After reviewing data from the BLS, IMF, WEF, Goldman Sachs, ILO, and a decade of academic research, the evidence supports the following conclusions:
1. Mass unemployment from AI is not supported by current data or credible near-term projections. Goldman Sachs found “no discernible impact” on labour market indicators as of early 2025. BLS projects 3.1% total employment growth from 2024 to 2034. The WEF projects a net gain of 78 million jobs by 2030.
2. Routine displacement is real, accelerating, and being systematically undercounted. Customer service, clerical, administrative, and routine knowledge-work jobs are genuinely contracting. The people in these roles are not going to automatically transition to being AI trainers. The skill gap is wide, the reskilling infrastructure is inadequate, and the timeline is compressed.
3. The ATM lesson applies — but with a critical speed caveat. Historical automation created new demand that offset displacement because the economy had decades to adjust. AI is moving faster. The Jevons Paradox may not have enough time to operate fully before a second wave of automation arrives.
4. The “AI creates jobs” and “AI destroys jobs” camps are both selectively reading the same evidence. The aggregate numbers look fine. The distributional picture is alarming. Both statements are simultaneously true.
5. The most important variable is not AI capability — it is reskilling velocity. WEF found that 59 out of every 100 workers globally will need reskilling by 2030. Whether they get it, at what cost, and who pays for it, will determine whether this transition is broadly prosperous or structurally cruel.
What This Actually Means — For Individuals and for Policy
If you are in a routine-heavy role — administrative, clerical, data entry, basic customer service, templated content production — the data is telling you something that your manager may not be saying out loud yet. The structural risk to those roles is not hypothetical; it is being priced into hiring decisions right now.
If you are in healthcare, skilled trades, complex technical work, or any role requiring judgment, physical adaptability, or genuine human connection, the same data says your employment outlook is strong and improving.
For policymakers, the numbers are pointing at a very specific failure mode: an economy where aggregate employment looks fine but the distribution of that employment becomes increasingly bimodal — highly paid knowledge workers and AI-adjacent roles on one end, low-wage service jobs that machines can’t yet do on the other, with the middle hollowing out.
That’s not a dystopia. It’s also not the “AI will create 78 million jobs, everything is fine” story either.
It’s a transition that will reward the prepared and punish the ignored — and the data is clear enough that we have no excuse for being surprised by where it goes.
Sources & References
- World Economic Forum — Future of Jobs Report 2025 (Survey of 1,000+ employers, 14M+ workers, 55 economies)
- IMF Blog — “AI Will Transform the Global Economy. Let’s Make Sure It Benefits Humanity.” — Georgieva, January 2024
- U.S. Bureau of Labor Statistics — 2023–33 and 2024–34 Occupational Employment Projections (Monthly Labor Review, August 2024 & November 2024)
- Goldman Sachs Research — “The Potentially Large Effects of Artificial Intelligence on Economic Growth” (March 2023); Goldman Sachs Labor Market Update (2025)
- BLS Monthly Labor Review — “Growth Trends for Selected Occupations Considered at Risk from Automation” (2022)
- BLS Monthly Labor Review — “Incorporating AI Impacts in BLS Employment Projections: Occupational Case Studies” (2025)
- Acemoglu, D. — Nobel Laureate, MIT — 2024 paper on AI productivity impacts; interview with Goldman Sachs Top of Mind series (June 2024)
- Bessen, J. — “Toil and Technology,” IMF Finance & Development (March 2015) — ATM/bank teller analysis
- International Labour Organization (ILO) — “Generative AI Likely to Augment Rather Than Destroy Jobs” (2023)
- PwC — 2025 AI Jobs Barometer — 25% wage premium for AI-skilled workers
- Dallas Federal Reserve — “Will AI Replace Your Job? Perhaps Not in the Next Decade” (2025)
- W.E. Upjohn Institute for Employment Research — “AI Exposure and the Future of Work” (2025)
- National University — “59 AI Job Statistics: Future of U.S. Jobs” (2026) — Aggregating BLS projection data
- Site Selection Group — Customer service employment decline data (2022–2024)
Data compiled and cross-referenced as of April 2026. All BLS projections are drawn from official publications and reflect the most recent projection cycle available at time of writing.