MIT Has Shown Us the Iceberg Beneath Microsoft’s X-Ray: What 151 Million Simulated Workers Reveal About AI’s Hidden Labor Market
With new data from MIT, we follow-up on our post titled Microsoft Research Just Mapped the Future of Work: What 200,000 AI Conversations Across 923 Jobs Reveal
TL;DR FAQ: MIT’s Project Iceberg Fills In What the Microsoft Study Couldn’t See
▼ Q: What is Project Iceberg?
A: A collaboration between MIT and Oak Ridge National Laboratory that simulated 151 million American workers as autonomous agents across 3,000 counties, mapping their skills against a catalog of over 13,000 AI tools. It produced the Iceberg Index, a new metric for measuring AI exposure before it shows up in unemployment data.
▼ Q: How is it different from the Microsoft study?
A: The Microsoft study showed us what AI is doing right now, based on real conversations. Project Iceberg simulates what AI can do technically across the entire workforce, including the parts of the economy that aren’t using AI yet. One is a diagnosis. The other is a forecast.
▼ Q: What is the “Iceberg” concept?
A: The visible tip is the 2.2% of labor market wage value currently affected by AI, concentrated in coastal tech hubs. The hidden mass below the waterline is 11.7%, five times larger, representing cognitive and administrative work across all 50 states that AI is technically capable of performing but hasn’t yet visibly disrupted.
▼ Q: Which jobs are in the hidden mass?
A: Administrative coordinators, financial analysts, HR professionals, customer support roles, insurance processors, healthcare schedulers, and white-collar workers in manufacturing supply chains. People whose jobs don’t look like “tech” but whose tasks are deeply cognitive and highly automatable.
▼ Q: What is “Automation Surprise” and which states face it most?
A: Automation Surprise describes the gap between a state’s visible tech-sector exposure and its true Iceberg exposure. Manufacturing states like Ohio, Michigan, and Tennessee show Surface Index values around 1 to 2% but Iceberg Index values approaching 12%, meaning ten times more of their workforce is technically exposed than current planning reflects.
▼ Q: Why don’t GDP and unemployment data reveal this?
A: Because those metrics measure what has already happened. The Iceberg Index found that traditional economic indicators like GDP, per-capita income, and unemployment explain less than 5% of the variation in AI skill exposure. Delaware and Utah show higher Iceberg exposure than California, despite much smaller economies.
▼ Q: What are states actually doing with this data in 2026?
A: Tennessee formally adopted the Iceberg Index in its 2026 AI Action Plan as the official framework for monitoring AI’s impact on its workforce. North Carolina and Utah are running similar analyses. The MIT team is expanding the Index to full national coverage at ZIP-code granularity so institutions can track exposure at the community level. In March 2026, Project Iceberg was also selected from over 800 applicants for the GitLab Foundation’s AI for Economic Opportunity Fund, backed by OpenAI, Ballmer Group, and the Anie E. Casey Foundation.
▼ Q: Does this mean mass layoffs are coming?
A: Not necessarily. The Index measures technical exposure, where AI can perform tasks, not displacement outcomes. Real-world impact depends on how firms adopt tools, how workers adapt, and what policy choices are made. Think of it less as a jobs countdown and more as a risk map.
▼ Q: What should employers do with this information?
A: Stop anchoring your workforce planning to what’s visible in tech headlines. Audit your administrative, financial, and coordination roles for task-level AI exposure. Develop reskilling pathways now, before adoption accelerates and the talent market hasn’t kept up.
▼ Q: What should job seekers do?
A: If you’re in a white-collar, cognitive, or administrative role anywhere in America, this research is about you, not just software engineers in San Francisco. The skills that protect you are the ones AI can’t replicate: complex judgment, stakeholder relationships, contextual reasoning, and domain expertise applied to novel problems.
▼ Q: What’s the bottom line?
A: Microsoft showed us the X-ray of how AI is being used today. MIT handed us the GPS for where it’s going. The five-fold gap between current adoption and technical capability represents both the biggest risk and the biggest preparation opportunity in the American labor market right now. And in 2026, the most forward-looking states are already navigating with it.
Late last year, we walked through what Microsoft Research found when it analyzed 200,000 real Bing Copilot conversations across 923 job types. That study gave us something rare and valuable: ground truth. Instead of predicting what AI might affect, it showed us what AI was affecting, in real workplaces, in real time. Information work dominated. Sales, service, writing, and analysis roles led the exposure rankings. Physically demanding jobs were largely insulated. It was a detailed X-ray of where AI has already landed.
That post generated more conversation than almost anything we’ve published. The question we kept hearing back was a fair one: okay, so that’s now, but what’s coming?
Enter MIT.
From X-Ray to GPS
Project Iceberg, a collaboration between MIT’s Media Lab and Oak Ridge National Laboratory, isn’t trying to observe current AI usage. It’s trying to simulate the entire technical landscape of where AI can go. The team built what they call a “digital twin” of the U.S. labor market, modeling 151 million workers as autonomous agents, each assigned their occupational skills, location, wages, and tasks. On the other side of the simulation, they cataloged more than 13,000 deployed AI tools, copilots, automation platforms, and workflow systems, and mapped those tools to the same skill taxonomy. Then they ran billions of interactions between the two populations on the Frontier supercomputer at Oak Ridge, one of the fastest in the world.
The result is the Iceberg Index: a skills-centered measure of the wage value of tasks that AI systems can technically perform within any given occupation. And what it reveals is, depending on your perspective, either reassuring or deeply unsettling.
Current visible AI adoption, the stuff in the tech headlines, the software engineers and data scientists and analysts using AI tools in coastal cities, represents about 2.2% of total labor market wage value. That’s approximately $211 billion. It feels enormous because it’s concentrated in high-profile industries that dominate the business press. But it’s the tip.
The Iceberg Index, measuring technical AI capability across the full workforce including all the cognitive and administrative work that doesn’t look like “tech,” comes in at 11.7%. That’s roughly $1.2 trillion in wage value. It’s geographically distributed across every state in the country. And it’s five times larger than what’s currently visible.
That gap is the story.
What the Microsoft Study Got Right and What It Couldn’t See
To be clear: the Microsoft study was excellent for what it was designed to do. It answered the question “what are real people actually using AI for right now?” with real data. No speculation.
The most common user goals in Copilot usage fell into four broad categories: learning, communicating, teaching and explaining, and writing, with learning and communicating being the most prominent. The AI, on the other side of those conversations, spent most of its energy in a service role, providing, explaining, teaching, and responding. A majority of AI-side activity was devoted to communicating and teaching or explaining information to the user.
At the occupation level, that translated into a clear pattern. The top occupations by AI applicability included Interpreters and Translators, Writers and Authors, Services Sales Representatives, Customer Service Representatives, Telemarketers, and Technical Writers, all roles deeply rooted in the creation, dissemination, and processing of information.
And just as importantly, the study identified a real floor. The groups with the lowest scores include occupations that require physically working with people, operating machinery, and other manual labor.
But here’s what the Microsoft study couldn’t capture by design: it only sees where AI is being used, not where it could be used. It’s bounded by current adoption. If a company hasn’t deployed AI tools, its workers don’t show up in the Copilot data. If an industry is slow to adopt, it looks unaffected. The surface looks calm.
That’s exactly the blind spot Project Iceberg was built to illuminate.
The Manufacturing Belt’s White-Collar Problem
The most striking finding in the MIT report is what they call “Automation Surprise,” the gap between how exposed a state looks on the surface and how exposed it actually is when you measure technical AI capability against the full workforce.
America’s industrial heartland shows the largest gaps. Rust Belt states such as Ohio, Michigan, and Tennessee register modest Surface Index values but substantial Iceberg Index values driven by cognitive work, including financial analysis, administrative coordination, and professional services, that supports manufacturing operations. Tennessee illustrates this pattern: a Surface Index of 1.3% but an Iceberg Index of 11.6%, indicating that administrative and service functions show up to ten times greater technical exposure than visible technology occupations.
Think about what that means practically. When policymakers in Ohio or Tennessee think about AI and work, they’re probably thinking about robots on the factory floor. Physical automation. The machines they can see. But the MIT data says the more immediate risk is the administrative and coordination infrastructure that keeps those factories running: the schedulers, the HR coordinators, the financial processors, the logistics planners, the customer service teams. That white-collar layer is deeply exposed to cognitive AI, and most of those states aren’t building preparation strategies for it.
Tennessee is the notable exception. The state formally adopted the Iceberg Index in its 2026 AI Action Plan as the official framework for monitoring AI’s workforce impact, citing the research directly. North Carolina and Utah are running similar analyses. It’s early, but it’s the right instinct: use the data before the disruption is visible, not after.
Meanwhile, technology-intensive states such as California and Washington show high values on both Surface and Iceberg Indices, resulting in smaller gaps. These states may recognize preparation needs earlier because skill overlap is already visible in their dominant technology sectors.
The states that most need to act are the ones least likely to see it coming. Tennessee just became the first to prove that doesn’t have to be the case.
The GDP Trap
One of the more intellectually uncomfortable findings in the MIT report is how badly traditional economic metrics fail to predict AI exposure.
Traditional economic metrics, including GDP, per-capita income, and unemployment, are widely used to benchmark state performance. Using 2025 data, the researchers ranked all 50 states by these conventional measures and then compared them to state rankings from the Iceberg Index. Their relationship with the Iceberg Index was negligible, with GDP, income, and unemployment each explaining less than five percent of the variation in systemic exposure. In some cases, the correlations were weakly negative.
Five percent. That’s not a weak correlation. That’s statistical noise. It means the economic data states currently use to make billion-dollar workforce investment decisions is nearly useless for predicting where AI exposure is actually concentrated.
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.
This has direct implications for talent strategy at the employer level too. Companies that benchmark AI readiness against their size, revenue, or industry reputation are flying blind. A mid-market insurance company in Sioux Falls may face more meaningful near-term AI exposure in its core workforce than a Fortune 500 manufacturer in Los Angeles. The metrics that make you feel big and safe have nothing to say about this.
The MIT team’s response to this gap is worth noting. Rather than simply publishing a paper, they built an interactive simulation environment that allows states to experiment with different policy levers, from shifting workforce dollars and adjusting training programs to modeling how changes in technology adoption might affect local employment and GDP. The Index isn’t just a diagnosis. It’s a planning tool. And in 2026, it’s being used as one.
Where This Lands for STEM Fields Specifically
In our last post, we walked through the sector-by-sector implications of the Microsoft findings. Project Iceberg adds a geographic and systemic dimension that changes some of those conclusions.
In tech: The Microsoft data showed high current exposure. The MIT data confirms it and extends it. Tech workers in coastal hubs are in the visible 2.2%. They’re already adapting. The more interesting talent planning question for tech companies isn’t whether their engineers will be affected. It’s whether the administrative and operational infrastructure around their engineers is being evaluated with the same rigor.
In manufacturing: The split we described in our last post, between AI-assisted CNC programmers and relatively insulated machine operators, remains valid. But the MIT data adds an entire hidden layer above the shop floor. The white-collar coordination roles in manufacturing, including procurement, quality documentation, logistics management, and HR, show significant Iceberg exposure in states like Ohio, Tennessee, and Michigan. These aren’t jobs that typically get treated as “AI jobs.” They are now.
In healthcare: Healthcare systems automate administrative tasks, enabling clinical staff to allocate more time to patient care. The MIT framing supports what the Microsoft data showed: clinical and hands-on care remains relatively insulated. But healthcare’s administrative infrastructure, including scheduling, billing, documentation, prior authorization, and patient coordination, is squarely in the cognitive automation zone. Healthcare employers who are only looking at clinical AI are missing the majority of their organizational exposure.
In life sciences: The research-facing roles remain protected by the complexity and novelty of discovery work. But the operational and regulatory infrastructure supporting research, including document processing, protocol administration, regulatory submissions, and literature synthesis, sits in the same cognitive automation zone as financial services and healthcare administration.
In sales: The Microsoft study already showed sales as the highest-exposure category by applicability score. The MIT data suggests the geographic spread of that exposure is broader than the current adoption data implies. Sales functions exist everywhere, including in manufacturing and healthcare organizations across the Manufacturing Belt. That exposure is just as real in Columbus as it is in San Francisco. It’s just not visible yet.
The 85% That Validates Both Studies
One of the most technically interesting parts of the MIT report is how they validated their framework. They had two tests: first, whether their skill-based occupational similarity scores matched actual observed career transition patterns. Second, whether their exposure predictions matched real-world AI adoption data.
Using skill-based embeddings derived from O*NET occupational profiles, they calculated similarity scores between all occupation pairs and compared these against career transition networks that capture observed worker mobility patterns. They found that 85% of commonly observed career moves involved occupations the framework identified as highly similar based on skills.
They also validated the Surface Index against the Anthropic Economic Index, actual AI usage data from millions of Claude users nationwide, and found 69% geographic agreement, with strong consensus at the extremes: 8 of 13 leading states and 9 of 13 aspiring states matched perfectly. Washington, California, and Colorado consistently appeared as leaders in both measures, while Wyoming, Mississippi, and Alaska aligned as laggards.
That cross-validation between the MIT simulation, the Microsoft observational data, and the Anthropic usage data is significant. Three different methodologies, three different data sources, arriving at a consistent picture of where AI exposure is concentrated today and projecting a consistent direction for where it’s spreading next.
What to Do With This: Practical Implications for Hiring and Career Planning
If you read our last post and updated your thinking about which roles are most affected, this follow-up should push you to update your thinking about which regions and organizational layers are most affected, and how far in advance of visible disruption you need to be acting.
For employers:
The Microsoft study said: break jobs into tasks, not titles. Project Iceberg adds: don’t just look at your tech roles. The cognitive and administrative infrastructure of your organization, regardless of your industry, likely has more AI exposure than you’ve estimated. Audit it. Build reskilling pathways for it before the adoption wave arrives and the talent market hasn’t caught up.
Employers in Manufacturing Belt states in particular should be running their workforce composition through a task-level AI exposure analysis now. Waiting for unemployment data to signal the problem is, per the MIT research, far too late. Tennessee just committed that approach to official state policy. That’s the bar.
For job seekers:
If you’re in a white-collar, cognitive, administrative, or coordination role in any industry in any state, Project Iceberg says the exposure is real, even if your company isn’t using AI tools yet. The question isn’t whether AI will reach your work. It’s whether you’re building the skills to work alongside it before it arrives.
The Microsoft data gave us the clearest practical advice on this: the work activities with the highest AI applicability and the highest scope scores are the ones that involve information creation, communication, and analysis. If you can demonstrate that you use AI to do those things better, not just that you’re aware of AI, you are positioned ahead of the curve.
As the Microsoft researchers put it: in the same way that word processors turned typing into a widespread skill rather than a core task for typists and secretaries, AI too may democratize skills around information work. That democratization is both a risk and an opportunity, depending on what you do with it.
The Summary That Both Reports Are Building Toward
Together, the Microsoft and MIT studies paint a picture that is more nuanced and more urgent than most AI-and-jobs coverage suggests.
The Microsoft study gave us the diagnosis: AI is already reshaping information work across nearly every sector, through a combination of task assistance and task delegation, with the greatest current impact in sales, service, content, and knowledge roles.
The MIT study gives us the prognosis: the transformation currently concentrated in coastal tech sectors is technically capable of spreading to five times that portion of the economy, through every state, into every industry with significant administrative and cognitive work, and it will arrive faster than traditional economic indicators will warn us.
In 2026, that prognosis is starting to become policy. States are adopting the Index. The MIT team is expanding it to ZIP-code granularity. The gap between what AI is doing today and what it can do technically is where preparation happens, or doesn’t.
At STEM Search Group, we’re not in the business of predicting exactly when that gap closes. But we are in the business of helping organizations and candidates position themselves in front of it. These two studies, read together, are the clearest picture we’ve seen yet of what “in front of it” means.
Sources:
- Tomlinson, K., Jaffe, S., Wang, W., Counts, S., & Suri, S. (2025). Working with AI: Measuring the Applicability of Generative AI to Occupations. arXiv:2507.07935.
- Chopra, A., Bhattacharya, S., Salvador, D., et al. (2025). Project Iceberg: The Iceberg Index, Measuring Skills-centered Exposure in the AI Economy. MIT / Oak Ridge National Laboratory.
- MIT Media Lab. (February 2026). The Labor Market Has an Iceberg Problem. media.mit.edu
- GitLab Foundation. (March 2026). AI for Economic Opportunity Fund: Grantee Cohort Announcement. MIT Media Lab / media.mit.edu