11.7% of Jobs Gone? What MIT's 'Iceberg Index' Actually Says

If you follow AI news at all, you saw the number: MIT study finds AI can already replace 11.7% of the U.S. workforce. It ran on CNBC, Fortune, and a hundred reposts, and it landed with that particular weight a round number from a famous institution carries. So I did the thing this blog is supposed to do: I read the actual paper. The study is real and the work is careful. The headline is doing something the study refuses to do.

Where the number comes from

The source is MIT’s Iceberg Index (Project Iceberg), out of the Media Lab, posted on arXiv in late 2025 (paper; project site). It is not a survey of employers and not a tally of layoffs that happened. It is a simulation, built in three steps:

1. Take O*NET, the U.S. government’s occupational database, and break 923 occupations into roughly 32,000 distinct skills.
2. Catalogue some 13,000 production AI tools and use language models to map which of those skills each tool can technically perform.
3. Weight every skill by its wage value, using Bureau of Labor Statistics pay data, and simulate 151 million workers across 3,000 counties.

The headline output: the Iceberg Index finds that 11.7% of total U.S. wage value — about $1.2 trillion — is “technically exposed” to today’s AI. They pair it with a “Surface Index” of 2.2%, roughly the slice visibly being adopted right now. Hence the metaphor: most of the exposure sits below the waterline, larger than the visible adoption suggests.

As a piece of measurement, this is legitimate and even elegant. The problem is entirely in the translation from what it measured to what people think it said.

Three things the headline quietly changed

1. It’s wage value, not workers. 11.7% is a share of dollars of wages tied to exposed skills, not 11.7% of people. “11.7% of the workforce” swaps a measure of money for a measure of humans, and they are not interchangeable — high-wage skills weigh more than low-wage ones, so the percentage of people affected could be quite different in either direction.

2. “Exposed” is not “replaced.” This is the one that matters, and the authors could not be clearer about it:

“The Index measures technical exposure — where AI capabilities and human skills overlap — not displacement outcomes.” It “does not predict job losses, adoption timelines, or net employment effects.”

Exposure means an AI tool can technically perform a skill. Whether it is cheaper, legal, reliable, or acceptable to actually hand that skill over is a completely separate question the study deliberately does not answer. Their own example: AI is technically “exposed” to diagnosing rare diseases, but liability and cost keep a human in the loop. Capability is not adoption, and adoption is not displacement.

3. It’s an upper bound, by design. To compute exposure, the model assumes a skill demonstrated in one occupation transfers to the same skill everywhere else — which, in their words, “establish[es] an upper bound on exposure.” So 11.7% is the optimistic-for-automation ceiling, not a central estimate. The realistic figure for actual displacement is some unknown fraction of it.

Put those together and the honest one-line summary is: today’s AI tools have technical overlap with skills accounting for up to ~12% of U.S. wage value — a ceiling on capability, not a forecast of jobs lost. Which is a genuinely interesting finding, and a far quieter one.

Who paid for it, and why that’s worth a footnote

Not a smoking gun, but the kind of thing this blog is built to notice. The Iceberg Index was supported by compute from Oak Ridge National Laboratory, a MIT Media Lab fellowship, and — relevantly — API credits from Anthropic, an AI vendor. And the “Surface Index” is validated in part against Anthropic’s own Economic Index. So an AI company’s resources and data help produce and validate a study about how capable AI is. That doesn’t make the work wrong; it does make it worth asking, as always, who benefits from me believing this? A larger “below the waterline” number is not bad for anyone selling what’s below the waterline.

So — is it believable?

Two different questions, two different answers.

As what it actually is: yes, reasonably. It’s a transparent academic exercise with explicit limitations the authors state plainly. As a map of where capability pressure is concentrated — administrative, financial, and professional services, spread across the whole country rather than just coastal tech hubs — it’s useful and probably directionally right.

As “AI can replace 12% of jobs right now”: no. That claim is a media compression the paper’s own text contradicts in writing. It survives because it confirms a mood, not because the evidence supports it — the same dynamic that sent the opposite viral stat (“95% of enterprise AI pilots fail”) around the internet a few months earlier. A nuanced finding gets flattened into a scary round number, and the round number outruns its caveats.

There’s also a deeper limit baked into the method itself, and it’s the reason I keep a separate post on the subject. The study can only see a job as the sum of its catalogued skills, because that’s the only form in which a job exists inside O*NET. Everything a worker does that never became a nameable, listed skill — judgement, context, trust, the tacit surround — contributes nothing to the exposure score. Measuring “the skills AI can do” and reporting it as “the jobs AI can do” is the doorman fallacy at population scale: precise measurement of the part you can see, silence about the part that may matter most.

Conclusion

Your instinct to look under the hood was the right one. The Iceberg Index is solid work; the headline is a load-bearing exaggeration the study explicitly declines to make. Read it as “here is where AI’s technical reach is largest” and it informs you. Read it as “12% of people are about to be replaced” and you’ve been told a story the data doesn’t support.

The calm reading is the accurate one: capability is real and growing, displacement is a separate and much slower thing, and the gap between them is exactly where the actual decisions — and the actual mistakes — get made.

Sources

1. MIT, “The Iceberg Index: Measuring Skills-centered Exposure in the AI Economy” (arXiv) — the primary paper; methodology, the “exposure not displacement” language, and the upper-bound assumption.
2. Project Iceberg, report PDF — MIT project site.
3. CNBC, “MIT study finds AI can already replace 11.7% of U.S. workforce.”
4. Fortune, “MIT report: AI can already replace nearly 12% of the U.S. workforce.”
5. Marketing AI Institute, “New Data Reveals 11.7% of the US Workforce Is Already Exposed to AI Automation” — useful breakdown of the exposure-vs-displacement distinction.