The Iceberg Index: Measuring Skills-centered Exposure in the AI Economy
B · Useful, limited MIT and Oak Ridge National Laboratory exposure index, widely reported in late 2025; the scale is large and limits are stated, but the mapping of 13,000 AI tools to skills is hard to check and the result measures technical exposure only.
Summary based on the full text·Original: about 35 minutes to read
In short
Project Iceberg (MIT, with Oak Ridge National Laboratory and state AI policy offices in North Carolina and Utah) builds a simulation of the US labour market and a skills-based exposure measure, the Iceberg Index. It estimates the wage value of skills that deployed AI tools can technically perform in each occupation. Visible AI adoption in tech accounts for 2.2% of wage value; technical capability across administrative, financial and professional services reaches 11.7%, about five times larger and spread across all states rather than coastal hubs.
Key numbers
- 11.7%share of US wage value (about $1.2 trillion) in skills current AI tools can technically perform
- 2.2%share of wage value (about $211 billion) in visible computing and technology adoption, the 'Surface Index'
- 151 millionworkers modelled as agents across 923 occupations, 32,000+ skills and 3,000 counties
- 69%state-level agreement between the Surface Index and actual usage in the Anthropic Economic Index
- less than 5%share of the Index's variation explained by GDP, income and unemployment
As reported by the source.
English·PDF·opens iceberg.mit.edu
What it reports
- Surface Index (tech-sector adoption): 2.2% of wage value, about $211 billion. Iceberg Index (cognitive automation capability): 11.7%, about $1.2 trillion.
- Current AI systems can technically perform about 16 percent of classified labour tasks.
- Exposure is not concentrated on the coasts: industrial states such as Tennessee (11.6%) and Ohio (11.8%) show substantial values; Delaware and Utah exceed California.
- GDP, income and unemployment explain less than 5% of the variation, so standard indicators miss it.
- The tool lets states run scenarios on training, incentives and regulation before spending.
How they got it
151 million workers across 923 occupations and 3,000 counties were modelled as agents with skills from standard taxonomies (32,000+ skills). Over 13,000 production AI tools were catalogued against the same taxonomy. Large Population Models built with AgentTorch, run on Oak Ridge's Frontier supercomputer, combine the two. Validation: skill similarity predicts observed career transitions, and the Surface Index agrees with Anthropic Economic Index usage in 69% of states.
Limits
- Measures technical exposure, not displacement, adoption timing or wages; the authors say so, though press coverage often read 11.7% as jobs at risk.
- Assumes skills transfer across occupations, which the authors call an upper bound.
- The tool catalogue and its skill mapping are not fully reproducible from the paper.
- Validation anchors on tech-sector usage, which may not reflect adoption in other sectors.
Why it matters for companies
The index points at back-office cognitive work (admin, finance, professional services) as the larger exposure pool, not the software teams where AI is most visible today. For workforce planning, the skills-level lens is more useful than job titles, because exposure varies within occupations. Treat the figures as a map of where to look, then measure adoption and outcomes in your own workflows.
Summary based on the full text. Read the original for the full argument and data.Open PDF on iceberg.mit.edu