What the Index Measures
The Opportunity Data AI Exposure Index measures how much an occupation's core work tasks overlap with current AI capabilities. It balances one exposure factor against two protective factors:
- Digital work intensity increases exposure. Computer-based, data-heavy work is the frontier of AI automation.
- Human contact intensity decreases exposure. Empathy, direct care, and constant interpersonal interaction are difficult for AI to replicate.
- Physical task intensity decreases exposure. Hands-on skill, physical presence, and responsibility for others' safety currently require a human body on site. Advances in AI-powered robotics may erode this protection over longer horizons, but the dimension remains a strong differentiator for near-term workforce planning (5-10 years).
A score of 0 means fully protected. A score of 1 means fully exposed. The index does not predict job loss. It measures task-level overlap with AI capabilities. Whether exposed occupations are augmented, restructured, or displaced depends on adoption speed, regulation, and organizational decisions the index cannot capture.
The Nine O*NET Variables
All variables are drawn from O*NET 31.0 (U.S. Department of Labor, August 2026). O*NET surveys incumbent workers and occupation analysts to produce standardized descriptors, and 911 occupations carry all nine variables in this release. Scores are self-reported survey responses, not modeled or imputed.
† Label correction (September 2026): earlier documentation listed this variable as "Contact With Others." The element identifier and every published score always used O*NET element 4.C.1.a.2.l, whose name is Face-to-Face Discussions; Contact With Others is a different element (4.C.1.a.4) that the index has never used. The September 2026 rebuild's variable-by-variable mapping check caught the mismatch, and reproduction of all 966 prior scores from archived O*NET 23.1 data confirmed no score was affected. The current version corrects the label and keeps the element.
Why "Responsible for Others' Health and Safety" is in the Physical dimension: This variable captures the requirement for physical presence. If someone's safety depends on you being there (a nurse, a firefighter, a pilot, a lifeguard), the work cannot be performed remotely or delegated to software. It complements the hands-on and posture variables by adding a safety-stakes signal that pure manual-work measures miss.
Three Dimensions
Digital Work Intensity
How computer- and data-intensive is the core work? Average of three normalized variables: Working with Computers, Analyzing Data or Information, Processing Information.
Physical Task Intensity
Is the work grounded in physical presence, manual skill, or safety responsibility? Average of three normalized variables: Spend Time Using Hands, Spend Time Sitting (inverted), Responsible for Others' Health and Safety.
Each dimension ranges from 0 to 1 after normalization. Digital intensity drives exposure up. Human contact and physical task intensity drive it down.
Construction: From Raw Scores to Composite Index
1
Extract and normalize. For each O*NET occupation, extract the raw score (1-5) for all 9 variables. Occupations missing any variable are excluded, leaving 911 scored occupations. Min-max normalize each variable to [0, 1] across the 911. Invert "Spend Time Sitting" so that higher values mean less sitting.
2
Compute dimension scores. Each dimension is the simple average of its constituent normalized variables.
Digital = mean( computers, analyzing, processing )
Contact = mean( caring, public, contact_with_others )
Physical = mean( hands, 1 − sitting, responsible )
3
Compute the composite. One exposure factor minus two protective factors, rescaled to [0, 1].
AI Exposure Index = ( Digital − Contact − Physical + 2 ) / 3
The +2 offset ensures the result stays in [0, 1]. All three dimensions receive equal weight (1:1:1).
Occupation-Level Index
696 occupations covering 119.5 million workers
The occupation index scores each Standard Occupational Classification (SOC) code directly from O*NET variables and pairs it with national employment estimates from the BLS Occupational Employment and Wage Statistics (May 2025). Of the 911 scored O*NET occupations, 696 map to unique 6-digit BLS codes. The remaining 215 are O*NET-specific subdivisions (e.g., "Data Warehousing Specialists") that map to broader BLS categories.
Each occupation receives scores on all three dimensions plus the composite index, alongside its BLS employment count. This is the foundation layer: every program-level score traces back to these occupation scores.
Academic Program-Level Index
1,806 academic programs across the full federal CIP taxonomy
The program index maps each Classification of Instructional Programs (CIP) code to occupations via the NCES CIP2020-SOC2018 Crosswalk. For each program, the linked occupations' AI Exposure scores are averaged using BLS national employment as weights. Programs whose linked occupations all have zero BLS employment receive an unweighted average.
Employment weighting
Employment weighting ensures that a program's score reflects the occupations graduates actually enter, not just the full list of possible jobs. Larger occupations contribute proportionally more to the program score.
Example: CIP 51.3801 (Registered Nursing) links to SOC codes for Registered Nurses (3.2M employed), Nurse Practitioners (355K), Nurse Anesthetists (45K), and others. The program's score is the employment-weighted average of those occupations' individual scores, so the 3.2 million Registered Nurses dominate the program's composite.
Data Sources and Update Schedule
Three federal inputs, one table: what the index uses, and when each input triggers a rebuild. Each rebuild updates this table and the vintage notes on every index page.
The state portfolio analysis additionally uses NCES IPEDS completions, which are released annually; state scores are updated when a new award year becomes final. BLS Employment Projections, released each year in late August, are not an input to the index: projections estimate expected employment growth, while the index measures task overlap with current AI capability. The two are designed to be read side by side, not substituted for each other.
Data refreshes versus index versions
A data refresh recomputes all scores on newer vintages of the same nine variables, using the same construction. Refreshes keep the methodology version (currently v2) and update only the vintage notes. A new index version is published when the construction itself changes: variables added or removed, dimension weights revised, or the aggregation rule altered. If O*NET or another federal source begins publishing occupational measures built specifically around AI exposure, those measures would enter through a new version released alongside the current one, with a change note documenting what moved and why. Prior versions and prior vintages are archived and available on request, so any published figure can be traced to the exact release it was computed from.
Status: updated September 2026 to O*NET 31.0 and OEWS May 2025; methodology unchanged (v2). The new O*NET uses the 2019 occupation taxonomy (the old data used 2010), so scores are re-baselined rather than comparable as a time series; the
update notes document the bridge and one label correction, and the
old version remains available. Separately, BLS published
AI Exposure Categories with the 2025-35 employment projections on August 27, 2026: four ordinal exposure levels for 831 occupations, synthesized from five external research and usage sources. The BLS categories are not an input to this index; the
comparison analysis measures where the two agree and diverge.
Limitations and Scope
Design assumption: frontier sufficiency. The index does not measure AI capability and does not age with it. It treats current frontier capability as sufficient for digitally mediated tasks and measures what binds when capability does not: hands-on work, physical presence, face-to-face interaction, and responsibility for others' safety. The assumption has two stated boundaries, both monitored: the physical anchor holds in the absence of capable AI-powered robotics, and the human-contact dimension reflects institutional and human preference for people in the loop rather than a technical limit.
What the index does
- Measures how much an occupation's core work tasks overlap with current AI capabilities, based on standardized federal survey data.
- Provides a consistent, reproducible score for every occupation in O*NET and every academic program in the federal CIP taxonomy.
- Identifies three distinct dimensions of AI exposure, allowing users to see why a score is high or low.
What it does not do
- Predict job loss. High exposure means task overlap with AI, not inevitable displacement. Many exposed occupations will be augmented rather than replaced.
- Account for adoption speed. Regulatory, organizational, and economic factors determine how quickly AI is deployed in a given field. The index measures technical susceptibility only.
- Reflect individual variation. A single score per SOC code. A software engineer at a startup and one at a defense contractor have different work contexts but the same score.
- Track AI progress in real time. O*NET 31.0 was published in August 2026, but descriptor ratings are collected on a rolling cycle, so an occupation's ratings can predate the current generation of AI systems by several years. AI capabilities are evolving faster than any survey update cycle.
Known data limitations
- Self-reported survey data. O*NET scores reflect incumbent worker perceptions, not objective measurement of task content.
- Employment is a point estimate. BLS OEWS May 2025 is a single cross-section. Occupations growing or declining rapidly may be over- or under-represented in the weighting.
- O*NET-to-BLS mismatch. 194 O*NET occupations lack BLS employment data because O*NET uses finer subdivisions. These occupations are scored but show zero employment.
- Crosswalk is many-to-many. A single CIP code can link to dozens of SOC codes, and vice versa. The employment-weighted average is a reasonable summary but masks within-program variation.
- Equal dimension weighting. The 1:1:1 weighting across Digital, Contact, and Physical is a modeling choice, not an empirical finding. Different weights would produce different rankings.
Opportunity Data · AI Exposure Index · Methodology
Released April 2026 · Updated October 2026