Methodology

• • Every evidence item stores its own origin, publication date, evidence type, confidence, direction and magnitude.
• • Confidence reflects study design: randomised and production deployments (80–95), platform-level observational studies (65–80), cross-country official analysis and scenario estimates (55–75), vendor and survey claims (50–70).
• • Counter-evidence is stored and used. Studies that find limits to AI effects reduce components rather than being discarded.
• • Duplicate events are deduplicated rather than counted twice.
• • Only titles, dates and short summaries are stored from external feeds. Full third-party articles are never republished.

The baseline corpus currently contains 149 evidence rows. Feed inventories and collection details are proprietary and not published.

Every score row stores the score version, methodology version and an input hash derived from its component inputs and the weights used. Snapshots are append-only: the score_snapshots table is written once per occupation per day per version and never updated in place. If the methodology changes from v1.0 to a new version, historical values stay as they were.

Historical charts show a backcast: the current methodology applied at quarterly checkpoints using only the models and evidence that had been published by each date. This is a reconstruction, not an observation, and is labelled as such wherever it appears. Observed daily snapshots accumulate from the date the index started running (8452 recorded so far).

Headline everywhere is exposure_score (labeled “AI Exposure”).

published = round( blended ), where blended is the weight-normalised mean of capability (30%), task automation potential (30%), observed adoption (20%), labour-market signal (10%) and evidence strength (10%). The score is never deflated for thin evidence - instead scores with fewer than 4 evidence items carry a Low confidence badge. Task exposure edges currently reuse the shared v1 editorial baseline per task - they are not per-occupation measurements until marked curated. Occupations flagged hasCuratedTasks=false show core tasks only, carry a “Limited task data” notice, and are noindex + excluded from the sitemap.

• • O*NET 31.0 Database - occupation titles, SOC codes, task statements with importance/relevance ratings (Task Type Core/Supplemental) and work activities. Contains information from O*NET OnLine by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA), used under CC BY 4.0. Task statements are being imported per occupation (replacing the shared v1 baseline) and rated individually for AI exposure.
• • BLS Occupational Employment and Wage Statistics (OES) - employment and median wage reference values (approx, annual). Shown with “≈” and “approx. BLS OES” everywhere.
• • O*NET-SOC taxonomy - occupation titles, SOC codes, categories and task statements. Each job page links its O*NET record and shows a “Last updated” date from the score row.
• • Frontier model announcements - capability releases tracked as they are published.
• • Automation tool launches - agent frameworks and production tooling.
• • Macro labor context (FRED, BLS projections) - official economic series used as background, not as AI signals.
• • Empirical research corpus - causal and observational studies on AI and work, each with its own URL, date, confidence and direction.

Collection schedules, feed inventories and polling infrastructure are proprietary and intentionally not published - they are the part of this project that would make it trivially clonable. What's public is the output: scores, verdicts and the live radar.

• • The occupation set is a curated subset of O*NET-SOC, not the full taxonomy. Small and emerging occupations are under-represented.
• • Task-level exposure values are currently a shared editorial baseline per task, not per-occupation measurements. Secondary edges are hidden from display until curated; occupations with filler tasks are flagged hasCuratedTasks=false.
• • Capability ratings reflect provider claims and public benchmarks, not independent testing by this project.
• • Observed adoption data over-represents digital and technical occupations, because usage data comes from AI products whose user base is not representative.
• • Labour-market outcomes are mostly platform-specific and US-centred. There is little causal evidence at occupation level, and this index does not manufacture any.
• • Region is not currently modelled. Exposure varies substantially by country and this index does not claim to describe labour markets outside the United States.
• • Employment and wage values are approximate reference figures and are flagged as such everywhere they appear.