The Data Behind the Score: 12 Institutions, One Index

What It Provides

The O*NET (Occupational Information Network) is the most comprehensive occupational database in the world. It contains detailed profiles for 997+ occupations including task descriptions, required skills, abilities, work activities, work context, and education requirements.

How We Use It

We use O*NET as our foundational layer. Every occupation in our index maps to an O*NET-SOC code. Task descriptions feed our Task Automation dimension — we map each of the 20,000+ individual tasks against AI capability benchmarks. Physical demands ratings feed our Physical Requirement dimension. Skill requirements help calibrate Creativity and Social Intelligence scores.

What It Provides

The BLS Occupational Employment and Wage Statistics (OES) program produces employment and wage estimates annually for over 800 occupations. The BLS also publishes the Occupational Outlook Handbook with 10-year employment projections.

How We Use It

BLS data enriches every individual report with real-time labor market context. When you search for a job, we pull current employment counts, median/mean wages, and growth projections. This data doesn't directly feed the composite score — it provides context about labor market size and trajectory.

What It Provides

10-year employment projections for 800+ occupations, including projected job openings, growth rates, and industry shifts. Published biennially as part of the Employment Projections program.

How We Use It

Growth projections help contextualize risk — an occupation can score high on AI risk but still show employment growth due to expanding demand. We present these projections alongside AI risk scores to give a more nuanced picture in individual reports.

What It Provides

Carl Benedikt Frey and Michael Osborne's seminal 2013 study 'The Future of Employment' estimated automation probabilities for 702 U.S. occupations. Their methodology used machine learning expert classifications of 70 occupations to train a model that predicted automation probability for the remaining 632.

How We Use It

Their automation probabilities serve as calibration anchors for our Task Automation dimension. While our scoring system uses 6 dimensions (not just one probability), we validate our composite scores against Frey & Osborne's estimates — occupations they identified as highly automatable should score high in our index too. Significant deviations trigger manual review.

What It Provides

This paper analyzed which occupations are most exposed to GPT-family models by mapping AI capabilities to specific work tasks. Rather than broad automation probability, it assessed LLM exposure at the task level — distinguishing between tasks GPTs can do directly vs. with software built on top of them.

How We Use It

This is our primary input for the Cognitive AI Exposure dimension. The paper's task-level exposure scores directly map to our cognitive assessment. Jobs with high LLM task exposure score high on Cognitive Exposure regardless of their physical or regulatory characteristics.

What It Provides

Ed Felten's index maps 10 categories of AI capabilities (image recognition, language modeling, reading comprehension, etc.) to occupational ability requirements from O*NET. It produces an exposure score based on which AI capabilities are relevant to each job's required abilities.

How We Use It

We use Felten's index as cross-validation for our Task Automation and Cognitive Exposure scores. His capability-to-ability mapping provides an independent methodology to check our scoring. Where his index and our scores significantly diverge, we investigate whether our dimensional analysis missed something.

What It Provides

Surveys 800+ companies across 46 countries about expected changes in job roles, skills demand, and technology adoption over the next 5 years. Includes employer estimates of job creation, displacement, and skills evolution.

How We Use It

WEF data feeds our understanding of which skills are gaining vs. losing value. Their employer surveys help calibrate our Creativity and Social Intelligence dimensions — if employers report that creative and interpersonal skills are becoming more valuable, that validates our resilience scoring for those dimensions.

What It Provides

McKinsey's automation research analyzes 800+ occupations across 46 countries at the activity level (not just the job level). They break each job into component activities and assess which activities can be automated with current or near-term technology.

How We Use It

McKinsey's activity-level analysis provides granular validation for our Task Automation scoring. Their distinction between technical automability and actual adoption speed informs how we weight near-term vs. long-term risk. Their cross-country data helps us understand how U.S.-specific our scoring might be.

What It Provides

The OECD's annual analysis covers G20+ labor markets with focus on automation risk, skills policy, and employment trends. Their methodology assesses risk at the task level within occupations, often finding lower automation estimates than Frey & Osborne because they account for task heterogeneity within jobs.

How We Use It

OECD data provides important methodological calibration. Their finding that most jobs are partially rather than fully automatable informs how we construct our composite — we don't treat automation as binary. Their regulatory analysis feeds our Regulatory Barriers dimension.

What It Provides

The International Labour Organization's task index classifies 30,000 tasks across occupations with 50,000 human evaluations of automation susceptibility, cognitive complexity, and physical requirements.

How We Use It

The ILO's massive task classification provides the deepest granular validation of our scoring. 50,000 human evaluations of task automability give us statistical confidence that our AI-based assessments align with expert human judgment. Where they disagree, human evaluations take precedence.

What It Provides

Stanford's Human-Centered AI Institute publishes an annual comprehensive report tracking AI technical performance, investment, adoption, policy, and societal impact. Includes benchmarks showing which cognitive tasks AI has surpassed human performance on.

How We Use It

HAI's capability benchmarks directly inform our Cognitive AI Exposure dimension. When their report shows AI surpassing humans on a new cognitive task (e.g., reading comprehension, legal reasoning), we re-evaluate which occupations use that task and adjust scores upward. This makes our index responsive to actual AI progress.

What It Provides

Brookings' analysis maps AI exposure to 36 million U.S. workers with metropolitan-level granularity. They focus on which demographics and regions face the highest exposure, including analysis by education level, race, gender, and metro area.

How We Use It

Brookings provides demographic and geographic context that enriches our reports. While our composite score is occupation-level, Brookings data helps us understand which populations within an occupation face higher or lower risk. Their Social Intelligence analysis also feeds our resilience scoring.

What It Provides

The IMF's 2024 analysis estimates that 40% of global employment is exposed to AI disruption across 108 countries. Advanced economies face higher exposure (60%) than emerging markets (26%) due to their larger share of cognitive-intensive occupations.

How We Use It

The IMF's global exposure estimates provide macro-level calibration. If 40% of jobs are significantly exposed globally, our scoring should reflect a similar proportion in the moderate-to-high risk range. Their country-level analysis helps contextualize U.S.-specific scores within a global framework.