MIT has created a real AI-labor measurement project, but it has not found that agents have already replaced 11.7% of U.S. workers. Project Iceberg’s Iceberg Index estimates the share of wage-valued work whose underlying skills and tasks current AI systems could technically perform. Its widely cited result is about 11.7% of modeled wage value, or roughly $1.2 trillion—not 11.7% unemployment.
A separate MIT Work Analytics Lab map released in June 2026 produces different scenario estimates, including about $1.4 trillion in wage-bill-equivalent exposure under a current-capability, full-adoption assumption. Both are exposure instruments, not counts of layoffs.
What MIT actually built
Project Iceberg, developed by the MIT Media Lab with Oak Ridge National Laboratory, combines a simulated U.S. labor market with profiles of AI capabilities. The model represents workers through occupations, skills, tasks, wages and locations, then estimates where AI tools overlap with that work. The project’s materials and report are available from MIT Media Lab and the Iceberg report.
The simulated labor market
- About 151 million U.S. workers
- More than 32,000 skills
- About 923 occupations
- Roughly 3,000 counties
- Approximately 13,000 AI tools in the simulation
The result is a skills-centered exposure measure. It can show that a task performed inside an occupation is technically exposed even when the occupation also requires judgment, physical work, communication, licensing or accountability.
#1 Best Overall
The Iceberg Index
The “iceberg” separates visible exposure—especially obvious technology-sector adoption—from a larger, less visible layer in administrative, financial, health-care and professional work. The metaphor does not say that every below-the-surface task will be automated. It says capability can exist before procurement, workflow redesign, regulation or staffing changes become visible in official statistics.
What the 11.7% figure means
| Measure | Modeled result | What it represents |
|---|---|---|
| Visible exposure | About 2.2% of wage value, approximately $211 billion | More concentrated, readily visible technology-sector exposure |
| Broader technical exposure | About 11.7% of wage value, approximately $1.2 trillion | Wage value associated with skills and tasks that AI capabilities could technically perform |
These figures are weighted by wage value rather than by a simple headcount. The $1.2 trillion is therefore not a forecast of lost pay, severance, government savings or corporate profit. It is an exposure-weighting measure in the model. The original research record is dated October 29, 2025 (arXiv).
Why this is not a jobs forecast
Four terms must be kept separate when reading any AI-labor statistic:
| Term | Meaning |
|---|---|
| Capability exposure | An AI system appears technically able to perform some work. |
| Adoption | An organization actually deploys that system. |
| Substitution | The system performs work previously assigned to a human. |
| Displacement | Workers lose employment, hours, pay or bargaining power as a result. |
The Iceberg Index primarily addresses the first category. It does not observe layoffs, unemployment, employer adoption or the speed of implementation. A task may be technically automatable yet uneconomic because of integration costs, error rates, data restrictions, liability, regulation, customer preferences or the time required for human checking.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchHow the later MIT exposure map differs
In June 2026, MIT’s Center for Transportation and Logistics and Work Analytics Lab launched a public U.S. AI Labor Exposure Map. It is related to Project Iceberg but is not a simple update of the same index.
- It uses Bureau of Labor Statistics May 2024 Occupational Employment and Wage Statistics data.
- It incorporates O*NET-derived occupation and task dimensions and MIT Work Analytics Lab task estimates.
- Its capability inputs include Anthropic usage evidence and the “GPTs are GPTs” framework.
- Its scenarios assume adoption and substitution to calculate wage-bill-equivalent exposure.
| Scenario | Estimate | Qualification |
|---|---|---|
| Current-capability scenario | About 18 million full-time-equivalent workers and $1.4 trillion in annual wage-bill value | Hypothetical full adoption and substitution of currently measured capabilities |
| Broader theoretical scenario | About 36 million FTEs and $2.9 trillion | Uses a wider theoretical capability framework, not observed losses |
Because the datasets, capability definitions and assumptions differ, the $1.4 trillion, $2.9 trillion and $1.2 trillion figures should not be added together or presented as a time series.
Which work is exposed?
The strongest conclusion is breadth, not a definitive list of occupations that will disappear. The models identify exposure in administrative work, finance, health care, professional services, software, document processing, routine analysis, quality control and logistics-related operations. MIT’s explanation emphasizes that AI often automates skills rather than whole jobs (MIT Media Lab).
For example, a financial analyst may have document-review and routine-evaluation tasks exposed while retaining responsibility for judgment, client communication, compliance and the consequences of a decision. Exposure can reduce task time, raise output expectations, change the skills employers seek or create demand for verification and integration without eliminating the occupation.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Why agents change the unit of work
An ordinary chatbot produces a response to a prompt. An agentic system can break a goal into subtasks, use software tools, work through a longer sequence, maintain state, iterate and recover from intermediate steps. That makes the relevant unit a delegated workflow rather than a single answer.
Longer workflows also create compositional failure risks: incorrect assumptions, poor tool selection, authentication problems, state-management errors, missed exceptions and small mistakes that compound. A production deployment therefore needs permission controls, audit trails, approval checkpoints, rollback procedures and a named person accountable for the outcome. Vendor examples, such as OpenAI’s report on Codex adoption, describe one company’s experience and are not a neutral measure of the whole economy (OpenAI).
Does the index prove mass unemployment?
No. Earlier MIT research using online vacancies found changes in requested skills at AI-exposed establishments—including reduced hiring for some non-AI positions and increased AI hiring—but no discernible aggregate employment or wage effect at the occupation or industry level during the study period (MIT Shaping Work). That evidence predates today’s newest agents, so it is not a final verdict. It does show why technical capability and realized labor outcomes must be measured separately.
What can workers do with this information?
A national exposure percentage cannot determine an individual’s job security. Workers can make it useful by mapping their own role into tasks:
Free tools Windows power users keep installed
One-click scans. No signup required.
- List recurring activities, including document handling, analysis, customer interaction, physical work, supervision and exception handling.
- Mark which activities require original judgment, relationship management, legal responsibility, physical presence or access to restricted systems.
- Track where tools are being piloted and whether they reduce time, increase volume or add checking work.
- Build skills in domain expertise, verification, communication, process ownership and safe use of AI tools.
- Keep evidence of improved outcomes—quality, turnaround time, error reduction or revenue—rather than treating tool familiarity alone as protection.
AI may augment a worker, allow less-experienced employees to handle harder tasks, increase workload expectations or create new roles in oversight, evaluation, integration and support. Outcomes can differ sharply by region because industry mix, employer size, infrastructure and training capacity differ.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How employers and policymakers should use the index
The index is best treated as an early-warning and planning instrument, not an automatic staffing plan. Responsible uses include:
- Prioritizing local training and reskilling investments.
- Identifying regions and workflows for measurement before making staffing decisions.
- Testing whether a task should be automated, augmented or retained.
- Designing human review, authorization and accountability requirements.
- Comparing actual deployment, correction time, staffing, wages and worker outcomes with modeled exposure.
MIT describes its simulations as a way to test interventions before committing substantial resources (MIT Media Lab). A policymaker that turns exposure into a layoff list is using the measure beyond what it establishes.
How to evaluate any AI-labor claim
- Ask whether the number counts workers, jobs, tasks, hours, skills, wage value or output.
- Check whether it measures capability, adoption, substitution or displacement.
- Identify the unit: occupation, task, workflow or firm.
- Read assumptions about adoption, human review, prices, regulation and physical constraints.
- Ask how success is defined: plausible output, production-quality output or output accepted without correction.
- Check whether residual human work—verification, exceptions, supervision and accountability—is modeled.
- Look for distributional results by geography, income, education, age and industry rather than relying on a national average.
- Distinguish vendor-derived capability or usage data from independent outcome measurement.
Limitations readers should keep in view
- Technical ability is not economic feasibility.
- Task exposure is not occupation elimination.
- Full-adoption scenarios are hypothetical.
- Results depend on which models, tools, error standards and human-review assumptions are included.
- Wage-bill-equivalent exposure is not worker income lost.
- Simulated workers represent statistical patterns, not each person’s actual workflow or employment result.
- The main estimates describe the U.S. labor market and should not be generalized automatically to other countries.
For personal-finance decisions, that last distinction matters: an exposure estimate can inform training, emergency savings and career planning, but it cannot supply a probability that a particular person will be laid off.
The Bottom Line
MIT has made AI’s potential reach across U.S. work more measurable. The Iceberg Index’s 11.7% and the later map’s $1.4 trillion and $2.9 trillion figures describe modeled capability or scenario exposure—not observed mass replacement. The practical question is which organizations adopt these systems, under what safeguards, and how the human work, pay and responsibility around them change.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
