Methodology · Detail
Coefficient notes
What is derived from public data, what is expert judgement, and where every adjustable cell sits between the two. This page is the audit trail for United States (and the same structure applies to every other country in the tool).
What's derived vs. what's hand-set
| Field | Source | Status |
|---|---|---|
automation_prob | Frey-Osborne (2013) layered with documented modern AI overlay | Derived |
automation_prob_legacy | Original Frey-Osborne 2013 (preserved when overridden) | Derived |
aioe_general, aioe_lm, aioe_ig | Felten-Raj-Seamans AI Occupational Exposure (2018, 2023 generative-AI update). Programmatic ingest joined by SOC code; surfaced as parallel input alongside automation_prob for cross-checking. US only at present (other countries pending SOC↔ISCO/ANZSCO/NOC crosswalks). | Derived |
aei_soc_major_pct | Anthropic Economic Index (release 2025-09-15, data window 2025-08-04..2025-08-11). Observed share of Claude.ai conversations classified into this occupation's SOC major group, in this country. Empirical near-term adoption signal. US only; other countries hold the country-level automation/augmentation collaboration split in meta.aei. | Derived (observation) |
wage_mean | National statistical agency (BLS, ONS, ABS, StatCan) | Derived |
employment_count | Same source as wage | Derived |
customer_industries[].share | Expert judgement, light cross-checks against industry trade reports | Estimated adjust → |
customer_industries[].ai_exposure | Expert judgement, anchored on the same scenario literature as automation probabilities | Estimated adjust → |
ai_demand_boost | Expert judgement; literature is largely silent on second-order demand benefits at occupation level | Estimated adjust → |
entry_barrier | Expert judgement, one-line rationale per occupation (training length, licensing, capital, hiring caps) | Estimated |
labour_inflow_rate, labour_inflow_lag_months | Judgement call built from Goldman Sachs (2023) and IMF (2024) displaceable-share ranges; same in every country | Estimated |
The three customer-demand estimates can be edited from the presets page. Try one of the built-in lenses ("AI hits a wall", "AI eats everything") or clone one to set your own values. The labour-inflow fields are fixed in the data for now.
Entry barriers: who can move in
Workers displaced by AI look for jobs AI can't easily do. Whether they can actually get one depends on how hard the occupation is to enter. entry_barrier scores that from 0 (open) to 1 (closed): how hard it is for a displaced mid-career worker to get in within about two years, counting training length, licensing, start-up capital and capped intakes. It is a judgement call, not a measured value, and each occupation carries a one-line note explaining its score.
- Nearly closed: dentist and pharmacist 0.95, solicitor or lawyer 0.9, civil engineer 0.85. Years of study plus registration.
- Apprenticeship or degree: registered nurse 0.75, electrician 0.65, plumber 0.6.
- Short training: hairdresser 0.3, real estate agent 0.15, personal trainer 0.1, fast food worker 0.05.
The barrier combines with the occupation's own automation probability: displaced workers don't move into jobs that are themselves being automated. See the formulas page for how the inflow rate of 0.4 was derived.
Modern AI overlay - the post-LLM adjustments
Frey & Osborne (2013) was published before the rise of generative AI. Some of their probabilities have moved substantially in light of post-2022 LLM and image-model capability gains. The simulator applies a documented overlay that:
- Lifts software developer (0.04 → 0.42), graphic designer (0.08 → 0.82), marketing manager (0.01 → 0.30), financial analyst (0.23 → 0.70).
- Lowers plumber (0.35 → 0.10), auto mechanic (0.59 → 0.18), carpenter (0.72 → 0.20). Frey-Osborne over-weighted dexterity-automation gains that didn't materialise.
- Holds truck drivers, fast food workers, registered nurses, dentists.
Each override carries a one-sentence rationale citing post-2022 evidence. When an override is in effect, the original Frey-Osborne number is preserved as a legacy field and shown in the simulator title strip alongside the modern figure ("modern overlay · F-O was 4%"), so the reader can see both.
Validation
Every Frey-Osborne legacy probability matches the published 2013 supplementary table to four decimal places. Wages match the respective national statistical agency's most recent published figures within tight tolerance. Customer-industry shares within an occupation always sum to 1.0 ± 0.02. Every scenario S-curve sits inside the parameter ranges anchored on IMF (2024) and Goldman (2023).
What's not yet modelled
- Geographic dispersion within a country. A national mean masks regional variation (San Francisco software developers vs. Phoenix; London vs. Sunderland).
- Endogenous occupation-to-occupation cascades. When software developer income drops, the simulator does not yet automatically resolve the downstream effect on real-estate agents whose customer base includes those developers. The Moretti multiplier and lag are applied as an aggregate drag on local-service occupations, and labour inflow moves displaced workers into easy-to-enter occupations in aggregate, but neither traces a specific occupation-to-occupation link.
- Policy interventions. No retraining subsidies, sectoral tariffs, or transfer-payment scenarios.
- Income-decile distribution within an occupation. A single mean wage hides considerable spread.
- Customer-industry shares from input-output tables. The single largest model-level uncertainty. Replacing expert judgement with derived BEA / ONS / ABS / StatCan values is the next planned data revision.