Science & Technology · AI & Computing · APRIL 11 2026
The Iceberg Index: What MIT's Groundbreaking Study Reveals About AI, Jobs, and the $1.2 Trillion Blind Spot
The world has been watching the wrong part of the iceberg. A landmark 2025 study shows that AI's true disruption to the global workforce is five times larger than the headlines suggest — and hiding in plain sight.
Every technology wave in history has arrived alongside a wave of anxiety about which workers it will displace. The printing press threatened scribes. The loom threatened weavers. The computer threatened clerks. Each time, the dominant narrative fixed itself on the most visible victims, and each time it missed the deeper transformation reshaping industries that nobody was watching. The current AI moment is no different — and a remarkable 2025 study from MIT and Oak Ridge National Laboratory makes that case with unprecedented precision.
The headlines have been consumed by a familiar story: programmers let go, junior developers struggling to find work, and tech giants quietly replacing human teams with AI tooling. That story is real. But according to the research behind the Iceberg Index, it represents barely one-fifth of what is actually unfolding beneath the surface of the global economy.
Why the Old Tools Cannot See the New Problem
Every major economic transition has, at some point, forced economists to build entirely new instruments of measurement. In the industrial era, "output per hour" had to be invented because nobody had a framework for capturing physical productivity at scale. In the internet era, GDP ran into an embarrassing paradox: it could measure the value of encyclopedias sold in bookshops, but had no mechanism to capture the value of Wikipedia, which replaced them for free. The US Bureau of Economic Analysis eventually had to build a separate accounting framework — the Digital Economy Satellite Account — just to quantify services that existed entirely outside conventional market transactions.
The AI era demands a similar overhaul. Intelligence is now a shared input between humans and machines. The tools we inherited — GDP, unemployment rates, per capita income — were simply never designed to measure that reality. Traditional economic metrics explain less than 5% of the variation in actual AI workforce exposure across US states, meaning the instruments governments use to allocate billions in reskilling and workforce investment are essentially pointing in the wrong direction.
Building the Iceberg Index: How MIT Mapped the Invisible
The research team at MIT, in collaboration with Oak Ridge National Laboratory and its Frontier supercomputer, did not set out to predict which jobs will disappear. They set out to do something more precise: measure where AI capabilities and human skills currently overlap, weighted by the economic value of the work involved.
Their starting point was a digital simulation of 151 million American workers spread across 923 distinct occupations and 3,000 counties. To map what each occupation actually requires, the researchers used O*NET — the US Department of Labor's exhaustive database that breaks down hundreds of professions into their specific component skills and tasks: things like analyzing data, critical thinking, interacting with computers, coordinating with others, or writing technical documentation. Each skill carries an importance rating and a difficulty level derived from surveys of real workers in those roles.
They then catalogued more than 13,000 real, production-ready AI tools already deployed inside companies — from coding assistants and document processors to financial analysis software and workflow automation platforms. Every AI tool was put through the same O*NET skill taxonomy used for human workers. This created two maps using a shared framework, enabling a genuine apples-to-apples comparison between human capabilities and AI capabilities for the first time.
"Basically, we are creating a digital twin for the U.S. labor market." — Prasanna Balaprakash, ORNL Director and Co-Leader of Project Iceberg
The output is a single number for each occupation — a percentage representing how much of the wage value inside that job AI can already technically perform. This design choice matters enormously. An accountant might spend 60% of their time on document processing and data entry, but that time may represent only 35% of their economic value if the remaining work — complex judgment, client strategy, regulatory interpretation — commands a far higher premium. Weighting by wage value means the Iceberg Index reflects actual economic exposure, not merely task volume.
The Iceberg: What's Visible vs. What's Hidden
The visible tip of the iceberg — layoffs and role shifts in tech, computing, and information technology — represents just 2.2% of total wage exposure, or about $211 billion. This is where virtually every policy paper, every worried editorial, and every corporate restructuring announcement has been focused for the past two years.
But when the same methodology is applied across the whole economy, technical capability extends far below the surface through cognitive automation spanning administrative, financial, and professional services, reaching 11.7% — approximately $1.2 trillion in wage value — with exposure distributed across all 50 states rather than confined to coastal tech hubs.
That is a fivefold gap. The anxiety that has dominated public conversation about AI for two years has been aimed at roughly one-fifth of the actual problem. The remaining four-fifths sit below the waterline, in industries and roles that have generated almost no anxious headlines.
Who Is Actually Exposed — And It's Not Who You Think
The popular image of AI job displacement features a software developer or a call centre worker. The actual profile of the most exposed worker looks rather different. According to a complementary Anthropic study tracking real AI usage in professional settings, the most exposed group earns 47% more on average than the least exposed, is nearly four times as likely to hold a graduate degree, and is 16 percentage points more likely to be female.
These are knowledge workers — people whose professional lives are built around reading, writing, analyzing, synthesising, and summarising information. By every conventional measure, they did exactly what society asked of them: they invested in advanced education, pursued professional careers, and built expertise in cognitive work. They are precisely the group AI's cognitive capabilities overlap with most directly.
Source: National University AI Job Statistics, May 2025
AI Technical Capability vs. Actual Current Usage
Blue = high exposure · Orange = moderate · Green = low. Sources: MIT Iceberg Index 2025; Anthropic AI usage data, August 2025.
The Gap Between Capability and Reality
There is an important distinction the Iceberg Index draws carefully: technical capability is not the same as actual deployment. The gap between what AI can theoretically do and what it is currently doing in practice remains substantial, and that gap is what keeps much of the exposure theoretical for now.
For workers in computer and mathematics occupations, AI is technically capable of handling around 94% of their tasks — yet observed professional usage currently sits at roughly 33%. Similar patterns appear in legal work, architecture and engineering, and financial services. The capability exists; adoption is being held back by regulation, integration challenges, organisational inertia, and the persistent requirement for human oversight of AI outputs.
These friction points are real, but they are not permanent. As AI tools mature and as organisations build the institutional muscle to deploy AI at scale, the gap between capability and reality will narrow. The Iceberg Index is best understood as an earthquake risk map, not a weather forecast — it identifies which structures sit on fault lines, not precisely when the tremors will arrive.
The Real-World Data: What Is Already Happening
Abstract capability numbers are one thing. The employment data accumulating since 2023 tells a more concrete story, and for certain groups it is already alarming.
| Indicator | Data Point | Source & Period |
|---|---|---|
| AI-attributed job losses | ~77,999 tech job losses directly attributed to AI in the first 6 months of 2025 alone (427 layoffs/day) | DemandSage / Challenger Data, 2025 |
| Young developer employment | Employment for software developers aged 22–25 declined ~20% from its late-2022 peak by September 2025 | Stanford Digital Economy Lab, 2025 |
| Entry-level job postings | Down ~35% since January 2023; down ~15% year-over-year in 2025 | Revelio Labs / CNBC, 2025 |
| Big Tech graduate hiring | Big Tech reduced new graduate hiring by 25% in 2024 vs. 2023 | SignalFire Research, 2024 |
| AI-exposed age 22–25 unemployment | Rose ~3 percentage points since early 2025, well above peers in less-exposed trades | Goldman Sachs Research, 2025 |
| Enterprise hiring cuts due to AI | 66% of enterprises are reducing entry-level hiring because AI now handles those tasks | SHRM / High5 Research, 2025 |
| AI code generation | AI systems now write over 1 billion lines of code daily — exceeding total human developer output | MIT Iceberg Index Report, 2025 |
| Roles changed or eliminated | 91% of enterprises report that roles have changed or been eliminated by automation | SHRM, 2025 |
| McKinsey automation potential | Today's technology could theoretically automate ~57% of current US work hours | McKinsey, late 2025 |
| WEF net jobs projection | 92 million roles displaced by 2030; 170 million new roles projected — net gain of 78 million | World Economic Forum, 2025 |
The Geography of Exposure: States Nobody Is Watching
Ask anyone which US states are most exposed to AI disruption, and the answers are predictable: California, Washington, New York — the tech-heavy coastal hubs. The Iceberg Index delivers a very different answer.
Delaware and Utah exhibit higher Iceberg exposure than California, despite much smaller economies, because their concentrated finance and administrative sectors present sharper automation targets than California's diversified workforce. Tennessee, North Carolina, and Utah have co-authored the study and already begun building state-level policy scenarios using the platform's findings.
| State | Surface Index | Iceberg Index | Exposure Gap | Primary Vulnerability |
|---|---|---|---|---|
| Washington | 4.2% | ~9% | Moderate | Tech (surface visible) |
| Virginia | 3.6% | ~10% | Elevated | Federal admin, defence tech |
| California | 3.0% | ~9% | Moderate | Diversified — exposure spreads thin |
| Utah | ~2.5% | ~12% | Very High | Finance, back-office services |
| North Carolina | ~2.0% | ~12% | Very High | Administrative, financial services |
| South Dakota | ~1.5% | ~12.5% | Extreme | Financial services concentration |
| Tennessee | 1.3% | 11.6% | Extreme (10×) | White-collar support for manufacturing base |
| Ohio / Michigan | ~1.8% | ~11% | Very High | White-collar admin behind industrial economy |
Tennessee is perhaps the starkest illustration of the measurement problem. Its tech sector exposure is just 1.3% — barely enough to register in any standard workforce planning model. Its Iceberg Index is 11.6%. The white-collar workforce keeping Tennessee's factories operational is ten times more exposed than the tech sector everyone has been watching. Ohio and Michigan follow the same pattern: states that have spent years preparing for robots to take over their factory floors are about to discover that the white-collar disruption arrives first.
The 30% Who Cannot Be Automated — And Why That Is Its Own Problem
About 30% of the US workforce has essentially zero exposure to AI disruption. This includes cooks, mechanics, plumbers, nurses, childcare workers, electricians, and bartenders — people doing physical, relational, hands-on work that no language model can replicate. On the surface, these workers appear safe. The deeper economic dynamics are more complicated.
In 1965, Princeton economist William Baumol noticed something odd about the performing arts. A string quartet performing Beethoven required four musicians and roughly 25 minutes — exactly as it had a century earlier. Nothing about the performance had become more efficient, yet the cost of staging that concert kept rising, dragged upward by wages increasing everywhere else as manufacturing and agriculture became dramatically cheaper and faster. — The origin of Baumol's Cost Disease, and why AI is about to make it much worse
This is Baumol's Cost Disease, and AI is about to accelerate it substantially. If cognitive and administrative work becomes dramatically more productive — a financial analyst compressing a day's work into an hour, a software engineer producing the output of three people — then the relative cost of every service AI cannot touch will keep rising. The nurse still needs the same time per patient. The plumber still has to be physically present. Their output does not scale with AI, so the relative price of their work keeps increasing, pulled upward by a productivity surge happening in every adjacent sector.
This matters because most AI-proof work is essential and price-inelastic. People do not pull children out of school because costs rise. They do not skip the plumber when a pipe bursts. Healthcare, education, elder care, and skilled trades are not discretionary. And most of them are either funded or heavily subsidised by governments that are already fiscally stretched. Workers safe from AI disruption may find themselves in industries that governments will increasingly struggle to fund — even as those industries become more expensive relative to the AI-augmented economy around them.
Sector-by-Sector Exposure: The Full Picture
| Sector | AI Technical Exposure | Roles Most at Risk | AI-Resistant Roles | Disruption Timeline |
|---|---|---|---|---|
| Tech / Computing | 94% (task capability) | Junior devs, QA, data entry, code review | Senior architects, ML researchers | Already underway |
| Finance & Banking | 70–80% | Analysts, loan processors, compliance officers | Relationship bankers, M&A advisors | 2–4 years |
| Legal Services | ~80% | Paralegals (80% risk by 2026), legal researchers | Trial lawyers, complex negotiators | 1–3 years |
| Human Resources | 85–90% | Recruitment screeners, benefits admin, payroll | Culture officers, complex negotiations | Already accelerating |
| Healthcare | Admin ~55%; Clinical ~15% | Medical transcription (99% automated), coding | Surgeons, nurses, therapists, GPs | Admin: now; Clinical: 5+ years |
| Media / Journalism | ~65% | Content writers, junior reporters | Investigative journalists, on-ground reporters | 2–4 years |
| Manufacturing | ~30% (by mid-2030s) | Assembly line, quality control, packaging | Skilled maintenance, bespoke fabrication | 3–8 years |
| Skilled Trades | <10% | None significantly | Plumbers, electricians, carpenters | Minimal foreseeable risk |
| Healthcare (clinical) | ~15% | Routine admin, medical coding | Nurse practitioners (52% growth to 2033) | Minimal clinical risk |
What the Numbers Cannot Tell You
The Iceberg Index is candid about its own limitations, and that candour is part of what makes it credible. It does not account for physical robotics — this version measures only digital and cognitive AI capabilities. It measures technical overlap only, not outcomes. It cannot predict when firms will act, how fast regulators will respond, or which workers will successfully retrain. It draws no conclusions about whether AI creates more jobs than it destroys over the long run — a question where the honest answer remains genuinely uncertain.
The broader macro data is more ambivalent than headline numbers suggest. The Federal Reserve Bank of Dallas, reviewing wage and employment data through early 2026, found that in jobs with significant AI exposure, wages were not uniformly declining, suggesting that for many workers AI is currently augmenting rather than replacing their output. Goldman Sachs estimates the transition-period unemployment impact at roughly 0.5 percentage points above trend — and historically temporary.
How We Got Here: A Timeline of AI's Labor Impact
For two years, governments, companies, and workers have been navigating one of history's most significant economic transitions using a decades-old map. The instruments they rely on — GDP, unemployment rates, sector employment figures — were built for a world where intelligence belonged exclusively to human workers. They have no mechanism for measuring a world where intelligence is increasingly shared between humans and machines, at scale, across every sector simultaneously.
The Iceberg Index is not a prophecy. It does not tell you that 11.7% of US workers will lose their jobs. It tells you that the economic value embedded in those roles is technically performable by AI tools that exist today, weighted by what that work is actually worth. That is a different, more precise, and considerably more useful piece of information than anything the existing toolkit provides.
The most dangerous response to this data is complacency in the wrong direction — either panic about a catastrophe that may unfold gradually enough to manage, or dismissal of a structural shift that the data already shows is real and accelerating. The smarter response is to take the map seriously, understand that the buildings sitting on the fault line are not the ones making headlines, and start planning before the tremors rather than after.
The Iceberg Index is a skills-centred metric developed by MIT and Oak Ridge National Laboratory that measures the wage value of skills AI systems can technically perform within each occupation. It simulates 151 million US workers interacting with 13,000+ real AI tools, using the O*NET skill taxonomy to create an apples-to-apples comparison between human and AI capabilities. It is not a job loss prediction — it measures technical capability overlap, weighted by economic value.
No. The index measures what AI can technically do today, not what firms will actually do with that capability, or when. Adoption is constrained by regulation, integration challenges, cost, and organisational inertia. The gap between technical capability and real-world deployment remains large. The index functions as an early warning map, not a countdown clock.
California's workforce is highly diversified, so AI exposure spreads across many sectors and dilutes. States like Tennessee, Utah, and South Dakota have economies heavily concentrated in administrative, financial, and back-office services — precisely the cognitive tasks AI handles most effectively. Tennessee's tech sector exposure is just 1.3%, but its Iceberg Index score is 11.6%, making its white-collar workforce ten times more exposed than the tech sector alone.
About 30% of the workforce has near-zero AI exposure. This includes skilled tradespeople (plumbers, electricians, carpenters — 94% of construction firms report worker shortages), clinical healthcare workers (nurse practitioners projected to grow 52% by 2033), and hands-on service roles like chefs, childcare workers, and mechanics. The caveat is Baumol's Cost Disease: even these safe workers may face rising costs in their sectors as the AI-augmented economy around them becomes dramatically cheaper and more productive.
The macro projections lean net positive: the WEF estimates 92 million roles displaced but 170 million new roles created by 2030 — a net gain of 78 million jobs. Goldman Sachs projects transition-period unemployment to rise only about 0.5 percentage points above trend, and historically such effects are temporary. However, the key challenge is distribution: new roles will not be in the same locations, industries, or skill categories as those displaced, and the transition period can produce significant hardship for affected workers even if the long-run aggregate looks positive.
This article is for informational and educational purposes. It synthesises publicly available research and does not constitute financial, career, or investment advice.