From Ritu Raj | Product & Market Analysis
MIT's Iceberg Index Says AI Can Do 11.7% of US Wage Value. What It Measures
On this page
The MIT and Oak Ridge Iceberg Index estimates that current AI tools could technically perform skills worth 11.7% of US wage value, about $1.2 trillion. Within a day of release that became a headline saying AI could already replace nearly 12% of the workforce. The MIT AI study says the opposite in plain words: it does not predict job losses, adoption timelines or net employment effects.
Key takeaways
- The 11.7% is a capability map, not a layoff count. The paper states that the index "does not predict job losses, adoption timelines, or net employment effects," and compares it to an earthquake risk zone.
- Capability is measured by tool availability. A skill counts as automatable if one of 13,000 catalogued AI tools, drawn from MCP servers, Zapier and the OpenTools directory, can perform it.
- Visible adoption is about one fifth of the headline. The paper's Surface Index puts AI overlap in computing occupations at 2.2% of wage value, about $211 billion across 1.9 million workers.
- Aggregate US labour data has not moved with the exposure map. Unemployment was 4.2% in September 2026, inside a 4.1% to 4.3% band since March. Yale's Budget Lab called AI disruption "largely speculative."
What the MIT Iceberg Index actually measures
Short answer: the Iceberg Index measures the share of each occupation's wage value tied to skills that existing AI tools can perform. It is a measure of technical overlap between what workers do and what software can do today. It says nothing about whether employers will adopt those tools, how fast, or whether jobs disappear as a result.
The paper is titled "The Iceberg Index: Measuring Skills-centered Exposure in the AI Economy". It was first posted to arXiv on 29 October 2025 and revised on 26 November 2025. The lead author is Ayush Chopra of MIT, with co-authors from Oak Ridge National Laboratory, the North Carolina General Assembly and Utah's Office of AI Policy.
The authors are unusually direct about scope. Section 4.2 says the index "does not predict job losses, adoption timelines, or net employment effects". It adds that actual impacts "depend on firm adoption strategies, worker adaptation, regulatory choices, societal acceptance and broader economic conditions". That is a full list of everything a job-loss number would need, and the paper declines all of it.
The paper describes the index as "a capability map, similar to how earthquake risk zones identify exposure without predicting when events occur". I think that is the most useful sentence in the whole document. A seismic hazard map tells a city where to spend on building codes. It does not tell residents to move out next Tuesday.
Read that way, the 11.7% figure is a planning input for state workforce agencies. That is who the authors built it for, and it shows in the co-author list.
How the Iceberg Index model is built, step by step
The method has four moving parts. Each one is defensible on its own, and each one adds a layer of judgement that the headline number hides.
Step 1: a skills taxonomy of 32,000 items. The model starts from O*NET, the US Department of Labor's occupational database, and BLS skill taxonomies. Appendix A lists 41 work activities, 35 skills and 33 knowledge areas rated across 923 occupations. The paper says it tracks more than 32,000 distinct skills in total.
Step 2: a catalogue of 13,000 production AI tools. The AI side is a catalogue of "over 13,000 production-ready AI tools". The sources are Model Context Protocol implementations, the Zapier automation platform and the OpenTools directory. MCP is the open standard that lets language models call external software, which is covered in the analysis of MCP as an agent interoperability standard.
This is the design choice that matters most. A skill counts as automatable "if tools exist that language models can use to perform relevant tasks." Existence in a directory is the test. Quality, reliability and cost do not enter.
Step 3: a language model decides what each tool can do. To match tools to skills, the authors "use in-context learning with large language models to infer which skills each tool can perform, based on its task descriptions and metadata". They calibrated on 600 occupations, held 300 back for validation, and manually reviewed the output.
That means a tool's marketing description, read by a model, partly decides whether a human skill is exposed. The paper's acknowledgements thank Anthropic for API credits. None of this is improper, but it is a long way from observed task completion.
Step 4: weighting by wages across 151 million workers. Each skill is weighted "by its relative importance, automatability score, and prevalence". Employment and wages come from BLS OEWS data on about 134 million wage and salary workers. Geography comes from the American Community Survey, covering 151 million workers across 3,000 counties. The simulation runs on Oak Ridge's Frontier supercomputer.
The readable paper does not publish the weighting formula or define the automatability score. You can follow the logic. You cannot reproduce the number.
Iceberg Index vs Surface Index: 11.7% against 2.2%
The paper's best idea is the contrast that gives it its name. It computes two numbers. The Surface Index covers overlap in computing and technology occupations, where AI adoption is already visible. The Iceberg Index covers all digital cognitive work.
The Surface Index is 2.2%, about $211 billion across 1.9 million workers. The Iceberg Index averages 11.7%. The paper calls the gap "fivefold," and argues that the hidden exposure sits in administrative, financial and professional services work.
Tennessee is the paper's showcase
Tennessee has a Surface Index of 1.3% and an Iceberg Index of 11.6%. The paper describes that as "up to ten times greater technical exposure" in administrative and service functions than in visible tech jobs. On the published figures the ratio is about 8.9. Fortune reported that Tennessee, North Carolina and Utah have started using the platform for state workforce planning.
Section 5.6 reports that GDP, income and unemployment each explain less than 5% of the variation in systemic exposure across states. The authors present that as proof the index adds information. It also means the index cannot be checked against the indicators most people would use to test it.
How the 11.7% AI job exposure figure became a job-loss number
Fortune's headline on 27 November 2025 read "MIT report: AI can already replace nearly 12% of the U.S. workforce". The body of the same article said the index "is not designed to forecast specific layoffs." The headline and the article disagree, and the headline is what travelled.
My view is that the paper invited some of this. Calling the 11.7% "five times larger" than visible disruption, and attaching a $1.2 trillion figure, gives readers a big countable object. Readers then do what readers do with big countable objects, which is treat them as a count of jobs.
| Reported as | What the paper actually measures | Why the difference matters |
|---|---|---|
| AI can replace 11.7% of workers | 11.7% of wage value sits in skills a tool can technically perform | A worker with 20% exposed skills still has 80% unexposed work |
| $1.2 trillion of jobs at risk | $1.2 trillion of wages attached to exposed skills | Wages attached to a task are not wages that vanish if the task is automated |
| AI can already do this | A tool exists in a directory that a model judged relevant | No test of quality, error rate or cost at production volume |
| A forecast for the next few years | A static baseline with no adoption model | The paper says adoption dynamics are "future work" |
"Perform" is not "replace"
Exposure is measured at the skill level, then rolled up. A financial analyst might have a chunk of document processing exposed and the rest of the role untouched. The paper says exactly this: "financial analysts will not disappear, but AI systems may demonstrate capability across significant portions of document-processing and routine analysis work".
Partial task automation usually changes what a job contains before it changes how many people hold it. Whether that turns into fewer hires is a separate question, and the evidence on it is covered in the analysis of conflicting entry-level hiring data.
The dollar figure does not reconcile on the page
The paper's introduction describes a US labour market of about $9.4 trillion in wage value. On that base, 11.7% is roughly $1.1 trillion, not $1.2 trillion. The $211 billion Surface figure implies a base nearer $9.6 trillion, and the $1.2 trillion implies about $10.3 trillion.
The paper does not show the calculation, so this may simply be two different wage totals. It is a small gap. It is also the most quoted number in the coverage, and nobody reporting it appears to have checked the arithmetic.
What payroll data shows against the AI wage impact estimate
If 11.7% of wage value were being displaced, the aggregate labour data would show strain by now. Nearly a year after publication, it shows a soft market rather than a breaking one.
The BLS reported US unemployment at 4.2% in September 2026, within a 4.1% to 4.3% range since March. Payrolls rose by just 29,000 that month, after an average monthly gain of 45,000 over the prior year. July was revised down to a loss of 10,000. That is a weak hiring market, and weakness has several plausible causes.
The Yale Budget Lab's reading
The most careful attempt to find AI in the occupational data comes from Yale's Budget Lab. It tracks how fast the mix of occupations has shifted since ChatGPT's release in November 2022, and how unemployment duration has moved for highly exposed workers. As Fortune reported in February 2026, the lab concluded that anxiety about AI's effect on today's labour market "remains largely speculative".
Executive director Martha Gimbel put it more bluntly: "If you think the AI apocalypse for the labor market is coming, it's not helpful to declare that it's here before it's here". I agree with her, and the Iceberg authors would too, on the evidence of their own Section 4.2.
None of this makes the Iceberg Index wrong. A capability map and a payroll survey measure different objects, and they are not supposed to agree in the short run. The comparison is useful because it shows the size of the gap between "could" and "did". The same gap between announced and measured AI job cuts is broken down in the comparison of Challenger layoff counts with BLS payroll data.
How the Iceberg Index compares with earlier AI job exposure studies
The Iceberg Index is the latest in a line of exposure studies, and it is the most conservative of the well-known ones. That fact rarely appeared in the coverage.
The 2023 OpenAI and University of Pennsylvania paper "GPTs are GPTs" estimated that about 80% of the US workforce could have at least 10% of their tasks affected by large language models. It put around 19% of workers at 50% or more of their tasks affected. Its test was whether a model could cut the time for a task by half without reducing quality.
| Study | Test for "exposed" | Headline result |
|---|---|---|
| GPTs are GPTs, 2023 | Model could halve task time at equal quality, judged by humans and GPT-4 | About 80% of workers have 10% or more of tasks exposed |
| Iceberg Index, 2025 | A production tool exists that a model can use for the skill | 11.7% of wage value exposed |
| Surface Index, 2025 | Same test, tech occupations only | 2.2% of wage value exposed |
| Where the older study is stronger | Task-level judgement rather than tool-directory matching | Captures partial exposure across almost every job |
Why the numbers differ so much
The two studies ask different questions. "GPTs are GPTs" counted any worker with a slice of exposed tasks. The Iceberg Index weights exposed skills by wage value and requires a specific deployed tool. Neither is a forecast, and the gap between 80% and 11.7% should warn you against treating either as one.
The useful conclusion is that headline exposure numbers depend almost entirely on the threshold the authors pick. Change the threshold and the number moves by a factor of several.
Where this reading of the MIT AI study is weakest
I have argued that the 11.7% was misread. The case for taking it more seriously is stronger than most sceptical coverage allows.
The index may understate what is coming
The paper measures production tools, not frontier model capability, and calls that "conservative estimates of immediate exposure." Tool directories grow every month. A capability map drawn in 2025 is a floor for 2027, not a ceiling. If you plan on the assumption that exposure stays at 11.7%, you are probably wrong in the other direction.
Calm aggregates can hide concentrated damage
A 4.2% unemployment rate is an average across 160 million-odd people. Concentrated losses in specific occupations and age groups can sit inside a stable aggregate for a long time. Yale's own trackers note an uptick in occupational differences between younger and older graduates. The payroll comparison in this post cannot rule out localised harm.
The validation is thin but not nothing
The authors checked their state rankings against the Anthropic Economic Index, which records how people use Claude. They report 69% geographic agreement, with 8 of 13 leading states and 9 of 13 lagging states matching. That is a categorical match against one vendor's usage data, not a correlation. Section 6.4 admits that "validation is correlational rather than causal." It is still more validation than most exposure studies publish.
How to use AI exposure numbers in a headcount plan
This post is written for founders and directors deciding whether a number like 11.7% should change a hiring plan. My answer is that it should change where you look, not how many people you hire.
| Your question | Use this | Do not use this |
|---|---|---|
| Which roles could change first? | Skill-level exposure, Iceberg style, applied to your own role descriptions | National job-loss headlines |
| Will we need fewer people next year? | Your own measured throughput after a tool is deployed | Any exposure index |
| Is AI already hitting our labour market? | BLS data for your occupations and region | Vendor surveys of AI adoption |
| How much time does a tool save? | Timed before and after comparisons on real work | Self-reported time savings |
The method is the valuable part, and you can copy it in an afternoon. List the skills each role actually uses. Mark which ones a tool you already pay for can perform today. Weight by how much of the week each skill takes. You will get a better number for your business than any national index can give you.
Then test the marked skills on real work before you change a role. Self-reported time savings are the weakest evidence in this field, a problem examined in the analysis of what AI time-savings claims actually measure.
Frequently asked questions
What is the MIT Iceberg Index?
The Iceberg Index is a measure built by MIT and Oak Ridge National Laboratory of how much US wage value sits in skills current AI tools can technically perform. It simulates 151 million workers across 923 occupations and about 3,000 counties. Its headline result is 11.7%, or about $1.2 trillion of wages. The authors state that it measures technical exposure, not job losses or adoption timelines.
Does the MIT study say AI will replace 11.7% of jobs?
No. The paper states that the index "does not predict job losses, adoption timelines, or net employment effects". The 11.7% is the share of wage value tied to skills an AI tool could perform, rolled up from individual skills. A worker can have some exposed skills and keep the job. The replacement framing came from news headlines, not from the research itself.
What is the difference between the Iceberg Index and the Surface Index?
The Surface Index covers only computing and technology occupations, where AI adoption is already visible, and comes to 2.2% of US wage value, about $211 billion across 1.9 million workers. The Iceberg Index applies the same test to all digital cognitive work and reaches 11.7%. The gap is the paper's main argument, that most technical exposure sits outside tech jobs.
How does the Iceberg Index decide a skill is exposed to AI?
A skill counts as automatable if a production AI tool exists that language models can use to perform it. The authors catalogued over 13,000 tools from MCP servers, Zapier and the OpenTools directory. A large language model then inferred which skills each tool covers from its descriptions, with manual review. Quality, cost and reliability in real deployments are not part of the test.
Has AI caused job losses in US payroll data yet?
Not at a scale visible in the aggregate figures. US unemployment was 4.2% in September 2026 and has stayed between 4.1% and 4.3% since March. Yale's Budget Lab described AI disruption to the labour market as "largely speculative" in early 2026. Hiring is weak, but weak hiring has several causes, and concentrated effects in specific groups cannot be ruled out.
Is the Iceberg Index reliable enough to plan with?
It is reliable as a map of where to look and unreliable as a count of anything. The weighting formula is not published, capability is judged by tool availability rather than tested performance, and validation is a categorical match against one vendor's usage data. Use the method on your own roles, and test exposed tasks on real work before changing headcount.
How to read the next AI jobs study in five minutes
Before you share or act on the next big exposure number, find one sentence in the paper: the definition of "exposed". If it says a tool exists, a model could help, or a task overlaps, you are reading a capability map. If it says a role was eliminated and counted in payroll records, you are reading an outcome.
This week, run the skill-level exercise on one team. Pick the role with the most routine document work, list its skills, and mark the ones a tool you already license can do. If the marked share is under a fifth, the 11.7% headline describes your business about as well as it describes the economy. Treat it as a direction to watch rather than a plan to execute. The wider question of why measured productivity has lagged these capability claims is taken up in the analysis of the AI productivity paradox.
Related on the labour data
For what announced AI layoffs look like against measured separations, read Challenger counts against BLS payroll data. For the hiring side, see why entry-level hiring data conflicts.
References
- Chopra, A., Bhattacharya, S., Salvador, D., Paul, A., Wright, T., Garg, A., Ahmad, F., Schwarze, A. C., Raskar, R. and Balaprakash, P., The Iceberg Index: Measuring Skills-centered Exposure in the AI Economy, arXiv 2510.25137, v1 29 October 2025, v2 26 November 2025. Used for all index figures, method, validation and stated limitations.
- Fortune, MIT report: AI can already replace nearly 12% of the U.S. workforce, 27 November 2025. Used for the headline framing and the states using the platform.
- Fortune, If AI is roiling the job market, the data isn't showing it, Yale Budget Lab report says, 2 February 2026. Used for the Yale Budget Lab findings and Martha Gimbel quotes.
- US Bureau of Labor Statistics, The Employment Situation, September 2026, 2 October 2026. Used for unemployment and payroll figures.
- Eloundou, T., Manning, S., Mishkin, P. and Rock, D., GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models, arXiv 2303.10130, 2023. Used for the earlier exposure estimates.
Weakest point in the source base: the Yale Budget Lab findings are cited through Fortune's reporting, because the Budget Lab's own pages could not be retrieved during research. The Iceberg Index weighting formula is not published, so every figure from it is taken on the authors' calculation.
Related reading