Everybody has an opinion. Almost nobody has data. This is 81 sourced data points on what AI is actually doing to employment — not predictions, not exposure scores, not hot takes. What the numbers say, where they come from, and the one critical variable that nobody has measured yet.
AI didn't start the jobs squeeze. It's accelerating a 50-year shift of income from labor to capital, and the people feeling it first are the ones just starting their careers.
The short version, before the 81 data points below:
AI didn't start this fight. It's accelerating a structural shift from labor income to capital income that's been running since 1973 — two years after Nixon ended the gold standard, a decade before the personal computer existed.
Every technology wave since has followed the same pattern: productivity goes up, labor's share goes down.
The bottom quarter of US households spend 95%+ of their income on necessities. Bank of America transaction data shows spending by the lower third of households shrank in 2025.
PIMCO (February 2026): "AI is not creating a new dynamic. It's intensifying a structural shift from labor income to capital income that has been running for decades."
The productivity-pay gap briefly closed twice: during the late 1990s dot-com boom and during the 2021-2023 post-COVID tight labor market. Both were driven by extremely tight labor markets forcing wage increases. Source: EPI
Internationally, the same decoupling occurred in two-thirds of OECD countries — but the gap is smaller in Germany and Scandinavia, where stronger union bargaining and labor regulation compressed it. EPI's assessment: the gap is "near-entirely driven by intentional policy decisions," not inevitable technological forces. Source: EPI, OECD
Forget the predictions. This is what's actually been measured:
AI isn't replacing entire occupations. It's hollowing out the entry level — cutting 22-27 year olds while keeping 30+ year olds to supervise AI output. The result: fewer on-ramps into knowledge work careers.
52,000 US tech workers were laid off in Q1 2026 alone, with March up 40% year-over-year, as companies explicitly shift budgets from headcount toward AI investment. Source: Challenger Gray, company announcements
| Sector | Net Impact | What's Happening |
|---|---|---|
| Tech / Software | Net negative | 89,251 cuts through July 2025. Q1 2026: 78,500 tech layoffs with ~48% AI-attributed. May 2026 wave: Meta 8,000 (10%), Cloudflare 1,100 (20%, explicit "AI revolution"), Freshworks 500 (11%, "over half our code is AI-generated"). But aggressively hiring senior AI engineers. |
| Healthcare | Net positive | AI acts as augmenter, not substitute. Radiology employs MORE radiologists than before AI (NVIDIA Jensen Huang, Adobe Summit 2026) — AI handles routine reads, total demand expanded. Health aide employment grew across all demographics. |
| Customer Service | Net negative | Entry-level down 11%. 74% of firms have rolled back AI chat agents due to poor results (Sinch survey 2026) — many displaced agents partially or fully rehired. Senior agents retained for complex cases. |
| Banking / Finance | Net negative | Standard Chartered: 7,000 corporate roles (15%) by 2030, CEO Bill Winters: "replacing lower-value human capital with capital." Back-office in India/Malaysia targeted. First major bank to confirm AI-driven structural cuts at this scale. |
| Industrial / Chemicals | Mixed | Dow: 4,500 jobs (8%) January 2026, press release cited "streamlining via AI and automation." First major non-tech industrial AI-citing layoff — signal AI displacement is spreading beyond knowledge work. |
| Consulting | Net positive | PwC UK launching 2,000 new AI/data specialist hires (£100M investment, April 2026). Accenture, Deloitte, PwC all expanding AI units. Titles: "AI Transformation Manager," "Generative AI Implementation Lead." Salaries $100K-$250K+. |
| Government | Mixed | 300K cuts but driven by DOGE mandates, not AI. |
AI Engineer is the #1 fastest-growing US job title, up 143% year-over-year. Source: LinkedIn
But new AI jobs pay $0-220K and are highly technical, with no direct retraining path for displaced routine white-collar workers. Data annotation jobs exist but are gig-based, lacking stability. The "prompt engineer" job title decreased 30% from 2024-2026, but prompt engineering skills in job postings tripled — the skill got absorbed into AI Engineer and Applied ML Engineer roles.
The academic framework economists keep coming back to is from MIT's Daron Acemoglu and Pascual Restrepo. They identify three forces operating simultaneously:
From 1947 to 1987, displacement (-0.48%/yr) and reinstatement (+0.47%/yr) roughly balanced, producing stable labor share and rapid wage growth. From 1987 to 2017, the balance broke: displacement accelerated to -0.7%/yr while reinstatement decelerated to +0.35%/yr. That gap is the "great decoupling" the data has been showing for 40 years. Sources: Acemoglu & Restrepo NBER WPs, multiple years
Two April 2026 papers (Gupta et al. on "Agentic Task Exposure"; Gao & Huang on "Bounded by Risk, Not Capability") independently land on the same finding: the jobs AI can't touch aren't the ones most people would guess.
Both papers also identify a "compliance premium": jobs that require a human to absorb legal liability or make ethical judgements retain higher human value because organizations keep humans in those loops regardless of AI capability. Sources: Gupta et al. arXiv 2026; Gao & Huang arXiv 2026
In February 2026, MIT economists Daron Acemoglu and David Autor co-authored a paper titled "Building Pro-Worker Artificial Intelligence" (NBER WP 34854). Notable because Autor is the labor-economics optimist who has long argued new tasks always emerge, and Acemoglu is the pessimist who has documented the recent imbalance. Their alignment is rare and significant.
Their joint diagnosis: AI development is currently too focused on automation over augmentation. Of five technology types (labor-augmenting, capital-augmenting, automating, expertise-leveling, new-task-creating), only new-task-creating technology is unambiguously pro-worker. The current AI investment trajectory is heavy on automation; market failures and "pro-automation ideology" in tech push it that way. They argue the path can be steered with policy — tax incentives for new-task AI, human-in-loop requirements, redirected public R&D.
The real question is which AI gets built, not whether AI will hurt workers. That decision is being made right now, by investors, founders, and policymakers.
The April 2026 BLS report had headline unemployment holding steady at 4.3% with 115,000 nonfarm payroll jobs added. Indeed Hiring Lab labeled it "low-hire, low-fire on lower ground." That headline is masking a real divide at the bottom of the career ladder:
The Harvard companion paper to Stanford's "Canaries in the Coal Mine" is the second empirical anchor: Hosseini and Lichtinger (Harvard, updated May 2026) used completely different data — 62 million workers, 285,000 firms, 2015-2025 résumé and posting records — and found junior employment falls roughly 9% at AI-adopting firms within six quarters, while senior employment continues to rise. They coined the term "seniority-biased technological change" — the phrase most likely to migrate from working paper into Treasury and OECD documents next.
Crucially, this is a hiring freeze, not a layoff wave. Separation rates for juniors at AI-adopting firms actually fell slightly. Companies aren't firing juniors; they've dramatically slowed hiring them. The phrase circulating among engineering leadership is "eating our seed corn." The tasks juniors used to do — first-pass research, basic coding, document review, ticket triage — are also how juniors became intermediates and then seniors.
Oliver Wyman's 2026 CEO Agenda surveyed CEOs behind 10% of global market capitalization. The headline finding: 43% of CEOs plan to cut junior roles in the next 1-2 years, up from 17% in 2025 — the largest year-over-year shift in the survey's history. The traditional corporate pyramid (many juniors feeding a small senior layer) is being replaced by a barbell or diamond: fewer juniors, more mid/senior operators, AI tooling beneath.
The counter-current is hiring at the top. ManpowerGroup's 2026 Talent Shortage Survey of 39,000 employers across 41 countries shows AI Model & Application Development (20%) and AI Literacy (19%) leading the global hardest-to-find skills list for the first time. Lightcast finds a 28% wage premium on AI-skilled postings (~$18,000/year). PwC pegs the premium on advanced AI skills as high as 56%. Sources: Oliver Wyman, ManpowerGroup, Lightcast, PwC
A year ago, executives hedged AI-motivated layoffs with euphemisms like "efficiency" and "restructuring." In 2026 they say it directly. Cloudflare cut roughly 1,100 jobs explicitly described as "made obsolete by AI" while reporting record revenue. Intuit announced a 17% workforce reduction (~3,000 employees) to refocus on AI. GM laid off 600+ salaried IT workers while openly hiring for "AI-native engineering." Block (Jack Dorsey) executed a 40% workforce reduction citing AI productivity gains.
Challenger, Gray & Christmas attributed approximately 49,000 layoffs to AI in the first four months of 2026 alone. Whether every case is causally clean is beside the point — the category itself has been legitimized. "AI made these roles obsolete" is now a sentence executives can say to shareholders without controversy. Even Jensen Huang's pushback (calling CEOs blaming AI for layoffs "a lazy excuse") proves the point: you don't argue with frames that don't exist yet.
Meanwhile the Magnificent Seven plan a combined $725 billion in AI capex in 2026, up 77% year-over-year, while cutting workers. David Shapiro calls this the "Great Capital Reallocation": harvesting payroll to buy compute.
The cognitive-automation story now has a parallel in industrial robotics. Hyundai committed to producing 30,000 Atlas humanoid robots annually by 2028, deploying 25,000 of them internally across Hyundai and Kia plants. Each unit costs ~$145,000. Production begins at the Savannah, Georgia Metaplant America. Boston Dynamics is training Atlas via GPU-parallel reinforcement learning with millions of simulated factory hours.
The Korean Metal Workers' Union has already blocked Atlas deployment without a formal labor-management agreement — expected to be the flashpoint of summer 2026 contract negotiations and the template for blue-collar AI bargaining globally. The IFR reports 542,000 industrial robots were installed worldwide in 2024, the fourth straight year above 500,000, with China accounting for more than half. The physical labor market is now also in play. Sources: Boston Dynamics/Hyundai CES 2026 keynote, IFR World Robotics 2025
The post-labor discourse has moved from niche to institutional in roughly 18 months. The signals stacking in 2026:
The deepest version of the story (Shapiro, May 2026): "Companies are buying senior judgment without funding junior development. The labor market is gradually losing the ability to reproduce itself. That's not a recession. It's a reproduction crisis."
This is already underway, assembling itself in pieces, while most of the commentariat stares at the 4.3% headline and calls it fine. Source: David Shapiro, "How the Post-Labor Economy Is Building Itself," May 2026
Robert Gordon analyzed 150 years of labor data. His conclusion: technology has never produced permanent mass unemployment. Vanguard's 130-year analysis agrees — AI is likely a general-purpose technology that lifts productivity over time, not a force that simply eliminates work. Sources: Robert Gordon, Vanguard (2026)
But "no permanent mass unemployment" glosses over what happened to individual workers, transition by transition:
| Transition | Scale | Duration | What Happened to Workers |
|---|---|---|---|
| Agriculture | 50% of workforce to <2% | 50-60 years | Created the industrial workforce. One farmer now feeds 160 vs 25 in 1900. Devastating for displaced farmers during transition. |
| ATMs | Bank tellers GREW | 25 years | 485K tellers (1985) to 527K (2002). ATMs made branches cheaper, banks opened 43% more branches. Textbook Jevons Paradox. |
| Ride-sharing | Decimated taxi medallions | 5-7 years | Total driver pool expanded (gig economy), but individual driver income fell. |
| Robotic Manufacturing | 3.3 workers per robot | Ongoing | Every 1 robot replaces 3.3 workers. Workers shifted to lower-paying local service jobs. Acemoglu & Restrepo (2020) |
| China Shock | 1M+ manufacturing jobs | 1991-2011 | Workers stayed in place. Wages depressed for a full decade. Communities saw permanent declines in housing, tax revenue, plus deaths of despair. Autor/Dorn/Hanson |
| Deindustrialization | Millions of factory workers | 1980-2000 | 5 years after job loss, displaced workers still earned 25% less than previous salary. Human capital completely devalued. 1993 BLS study |
The pattern is clear: aggregate employment recovers. Specific people get destroyed. Displaced workers take 1 month longer to find work, suffer 3% initial earnings loss, and see real earnings grow 10 percentage points less over the following decade. They delay home purchases and are less likely to marry. Source: Goldman Sachs, 40 years of individual-level data
Retraining programs? The Trade Adjustment Assistance program had "vanishingly small" effects on per capita income. Retraining success rates for middle-aged workers are historically very poor. Source: o3 research, multiple studies
Everybody says AI will be "bigger than the internet." The actual math says otherwise:
| Technology | Peak Annual GDP Boost | Duration | Source |
|---|---|---|---|
| Steam Power | ~0.41%/yr | ~100 years to peak | Crafts (2004) |
| Electrification | ~1.0%/yr | ~40 years | Multiple sources |
| Computing / ICT | ~1.6%/yr | ~25 years | BLS, Jorgenson |
| AI (Goldman, bullish) | ~0.7%/yr | ~10 years projected | Goldman Sachs |
| AI (Acemoglu, bearish) | ~0.1%/yr | ~10 years projected | Acemoglu (MIT) |
Goldman's most bullish estimate (7% over a decade) puts AI at PC-revolution scale, not electricity-revolution scale. For AI to match electrification, you'd need a 15-20%+ GDP boost over a decade. No serious estimate projects that.
But Goldman projects AI's impact arriving in 10 years vs 25-100 historically. The speed is unprecedented even if the magnitude isn't. Acemoglu's sub-1% is widely critiqued for only counting direct task automation, not new capabilities or Jevons effects.
The enormous spread between 0.1%/yr and 0.7%/yr isn't random disagreement. It maps directly to a single unmeasured variable: price elasticity of demand.
University of Chicago economist Alex Imas thinks every major AI jobs study is fundamentally flawed: OpenAI, Anthropic, McKinsey, all of them score how exposed a job task is to AI, then estimate displacement from that score alone.
Imas's verdict, in his own words: "Exposure alone is a completely meaningless tool for predicting displacement."
The missing variable is price elasticity of demand. When AI makes a service cheaper, does demand expand proportionally?
This single variable determines whether AI creates or destroys jobs in any field. And nobody has measured it.
Imas has called for a "Manhattan Project" to collect price elasticity data across industries. The data doesn't exist. Current elasticity data only covers supermarket commodities via grocery scanner partnerships. Zero data exists for white-collar services. Source: Imas, University of Chicago
The GDP estimate spread is a price elasticity proxy. High elasticity across industries = Goldman's 7% (more demand, more GDP). Low elasticity = Acemoglu's sub-1% (same demand, fewer workers, marginal GDP gain). This one variable is the entire ballgame.
Jamie Dimon (JPMorgan annual letter, 2026): AI will "cure some cancers, reduce accidental deaths," and eventually reduce the work week to 3.5 days. But also: "AI will definitely eliminate some jobs" and deployment could outpace worker adaptation. Source: JPMorgan 2026 Annual Shareholder Letter
Goldman Sachs: Net 16,000 US jobs lost per month from AI. But at the macro level, AI-driven job displacement has contributed "basically zero" to 2025 GDP. The displacement is real but economically invisible at national scale — for now. Source: Goldman Sachs / Morgan Stanley
Vanguard (130 years of data): AI is most likely a general-purpose technology that lifts productivity over time. Early adopters capture the most value. Source: Vanguard Research, 2026
PIMCO (February 2026): AI isn't creating a new dynamic. It's intensifying a structural shift from labor income to capital income running for decades. Source: PIMCO Economic Outlook
Emotion recognition in the workplace: banned. By August 2026, all employment-related AI becomes "high-risk" — requiring bias testing, data logs, and human oversight. Source: EU AI Act text
Sanders/Kelly proposed AI automation taxes in 2026. No major legislation has passed. The federal response to AI displacement remains largely nonexistent.
OpenAI pivoted from universal basic income advocacy to a "tech-indexed public wealth fund." Their own UBI study results: $1,000/month for 3 years led to participants working 1.3-1.4 hours less per week — a moderate decrease in labor supply. Source: OpenAI UBI study
Retraining displaced workers sounds like the obvious solution. History says otherwise. The Trade Adjustment Assistance program showed "vanishingly small" effects. Retraining success rates for middle-aged workers are historically poor. The new AI jobs are highly technical with no clear bridge from the jobs being eliminated.
Bureau of Labor Statistics (labor share data) · Piketty/Saez (income distribution research) · Economic Policy Institute (productivity-pay gap) · Goldman Sachs (displacement methodology, individual-level displaced worker study, GDP estimates) · Morgan Stanley (macro displacement assessment) · Challenger Gray & Christmas (AI-attributed job cuts) · Stanford HAI AI Index / ADP (entry-level employment data) · Robert Gordon (150 years of technology and employment) · Vanguard (130-year analysis of general-purpose technologies) · PIMCO (structural shift analysis, February 2026) · Daron Acemoglu, MIT (GDP estimates, robot displacement studies) · Acemoglu & Restrepo (robotic manufacturing displacement, 2020) · Autor, Dorn & Hanson (China Shock research) · Alex Imas, University of Chicago (price elasticity critique) · Crafts (historical GDP from steam power) · Federal Reserve Survey of Consumer Finances (stock ownership) · Moody's / Federal Reserve (consumer spending distribution) · Bank of America (transaction-level spending data) · Jamie Dimon / JPMorgan (2026 Annual Shareholder Letter) · LinkedIn (AI Engineer job growth) · EU AI Act (employment provisions) · OpenAI UBI Study (labor supply effects) · Sanders/Kelly (proposed automation tax legislation)
I dig into what the employment numbers actually show — not the scary headline, not the cheerleading. When the next wave of “AI killed X jobs” stories hits, you’ll have the real figures.
— Scott Covert, who actually built this. I work with these tools hands-on every day. If something here sparked a question, a project, or a “wait — can you build that for me?”, that's the best message I get. Reach me, pitch a collab, or book a consult.