An MIT research team has published the Iceberg Index, a comprehensive mapping of 923 occupations and 32,000 skills that reveals how AI capabilities overlap with the workforce in ways that extend far beyond headline automation concerns.

The Iceberg Index provides a searchable dashboard at iceberg MIT that maps how much of each occupation’s skills and tasks current AI systems can already perform. The research team, which includes affiliates of the MIT Media Lab, constructed a taxonomy of more than 32,000 individual skills drawn from occupation-based data and rated each skill individually on a scale measuring how much of that specific task current AI systems can accomplish today.

The core methodology assigns each of the 32,000 skills a rating on a 0 to 1 scale reflecting current AI capability at performing that exact task. The occupation-level exposure score is then determined by the maximum of all individual skill ratings within that occupation. This “max skill exposure” approach means that an occupation’s exposure is driven by whichever single skill within the job carries the highest AI capability rating — not an average across all tasks in that role.

This approach produces very different results than previous studies based on full-job automation thresholds. An occupation where AI can fully automate any one of its component skills receives maximum exposure regardless of how many other tasks the occupation requires that AI cannot yet do.

Reading the Index

The Index presents AI exposure using a color-coded scale reflecting the percentage of AI skill overlap for each occupation. Occupations with scores near 0 show minimal current AI exposure. Those approaching 1 reflect skills where AI can perform at or above human level on that specific task. The dashboard groups results in categories ranging from minimal exposure to the highest exposure levels.

What the data shows

The research reveals significant concentration of AI exposure within specific clusters of occupations. Jobs that share core skills in content generation, pattern recognition, structured data processing and code generation show the highest overlap with current AI capabilities. Occupations requiring complex interpersonal interaction, physical dexterity, or real-time physical world navigation present lower exposure rates, as the team notes that intelligence emerges from interactions among agents rather than from any single model acting alone.

The framework identifies what the researchers call a ripple effect in workforce exposure — the phenomenon where AI capability in one specific skill creates exposure across a broader set of jobs that share that skill, even when the overall occupation involves many other tasks AI cannot yet perform.

According to the MIT team’s research documentation, they need to measure how millions of human and AI decisions interact before those interactions create hidden shocks across the economy. This framing shapes the entire Iceberg Index approach: rather than asking “which jobs can AI replace,” the Index asks “how much of each job can AI already do, and where does that overlap create the most leverage?”

Underpinning research

The Iceberg Index draws on MIT’s earlier work in differentiable agent-based modelling, framework for learning in agent-based models, and automated sensitivity analysis of agent-based systems. The team published methodologies in AAMAS 2025 and 2024 proceedings and submitted papers on arXiv documenting approaches to population-scale simulation of human-AI coordination.

The research framework incorporates MIT Media Lab work on multi-agent systems and agentic coordination levels, examining how smart agents fail at coordination in ways analogous to humans, and what that implies for workforce-level AI exposure analysis. Large Population Models form the computational underpinning of the methodology, enabling analysis at the scale required for economy-wide coverage.

Methodological implications

The “AI automates skills, not jobs” framing is central to the Iceberg Index interpretation. Traditional labor metrics — GDP, unemployment, per-capita output — measure workers and jobs rather than skills and tasks. The researchers argue these metrics have not yet caught up to how AI actually affects work, creating a blind spot between what AI can do today and how those capabilities manifest in employment outcomes.

The MIT team proposes that workforce adaptation strategies must account for the specific skills within each occupation that AI can now perform, rather than treating occupations as monolithic units that are either automatable or not. The methodology also considers different levels of AI capability across skill categories — from content generation where current models perform well to complex physical coordination and social reasoning where AI capabilities remain more limited.

Sources Cited