What is the AI Work Index?

The AI Work Index scores 562 Singapore occupations and 88 modern roles for structural AI displacement pressure. The structural score combines a multi-source exposure layer, a human bottleneck layer, and a Singapore market-resilience layer. No LLM is used in the scoring pipeline.

How is the AI job risk score calculated?

Net displacement risk = exposure ensemble × (1 − human bottleneck) × market modifier. The exposure layer combines available matches from AIOE, Anthropic observed usage, Eloundou GPT exposure, and the ILO occupational exposure index. Human bottleneck comes from Pizzinelli theta and market resilience from Singapore MOM demand, wage, and labour-market signals.

Is the AI Work Index open source?

Yes. The entire scoring pipeline, data, and website are MIT licensed and available on GitHub at https://github.com/kirso/aiworkindex. Anyone can reproduce the results by running the deterministic scoring script.

About This Project

562 occupations · 88 roles · No LLM in scoring · MIT licensed

Structural AI pressure scores for Singapore. Not a prediction of job losses — a measure of how much current AI capabilities overlap with each job's tasks, adjusted for human bottlenecks and local demand.

Structural Score

Core model. Exposure × bottleneck × market modifier. Published as the primary dataset.

Labour Monitor

Quarterly MOM data. Vacancy rates, hiring, retrenchment. Cluster-level, not per-occupation.

Offset & Support

Separate support layers. Offset potential, transition pathways, SkillsFuture programmes, and scenario guidance. Useful context, not a forecast.

These three layers are separated deliberately in the product and data model. The structural score is published as its own dataset, the labour monitor and Singapore context ship as separate artifacts, and the offset/support layers stay out of the core score so they remain auditable decision support rather than being presented as measured fact.

This model measures one side of the equation

In the Acemoglu & Restrepo (2019) framework, AI's net impact = displacement − reinstatement. We measure displacement only. Scores likely overstate net risk for occupations where AI creates new work.

State of the science (early 2026)

Single exposure scores are poor unemployment predictors — ensembles do better (Frank et al., 2025)
No consensus on measurement — "still in the first inning" (Brookings/PIIE, 2026)
Entry-level workers face earliest pressure (Stanford DEL, 2025; Anthropic, 2026)

Model Card

Direct / Reproducible

Reliability-weighted 4-source exposure ensemble when matched (AIOE + Anthropic + Eloundou + ILO)
Theta complementarity scores (O*NET survey data)
Net risk formula (fully reproducible)
Official demand signals (SOL 2026, Jobs in Demand)

Estimated / Group-Level

Market resilience (group-level employment trends + occupation wage structure)
Crosswalk quality (US occupations mapped to SG)
Labour monitor (cluster-level, not occupation-level)
Observed-usage calibration (Anthropic usage, not universal AI adoption)
BLS convergent check (ρ = −0.14, broad directional check)

Synthetic / Illustrative

Modern role estimates (weighted SSOC priors + workflow/context adjustment)
Transition support (deterministic feasibility estimates + official programme infrastructure)
Offset potential (heuristic demand, redesign, and friction layer)
Outlook/scenario modelling (rule-based guidance, not prediction)
Seniority modifiers (research-grounded, not independently validated)

Still Limited

Occupation-level backtesting is still limited; current public validation remains cluster- and family-level, not occupation-level
Company-size modifiers (not part of the current structural model)
Causal displacement claims are out of scope (current evidence is correlational)
Occupation-level employment counts (not publicly released; requested from agencies)

Data Vintage

Wages

2024 MOM data

Demand Signals

SOL 2026 + Jobs in Demand 2025

Labour Market

Q4 2025 full

Model Version

V4.1 · 129 checks

Inspiration & How We Differ

Inspired by Andrej Karpathy's AI Job Exposure Map (March 2026) and Josh Kale's extended visualization, which score 342 US occupations using LLM-generated ratings (Gemini Flash, 0–10 scale).

What we do differently:

No LLM in scoring — we use deterministic transforms of published research and official data, not live model-generated ratings
Singapore-specific — SSOC occupational classification, MOM demand signals (SOL 2026, Jobs in Demand), Singapore labour market data
Three-layer structural score — exposure ensemble, human bottleneck, and market resilience are kept separate rather than hidden inside one opaque score
Externally cross-checked — cluster-level directional checks, BLS convergent evidence, and 56 internal structural checks
Seniority modifiers — research-grounded experience level adjustments (Stanford DEL, Anthropic 2026)
88 synthetic roles — modern job titles (AI Engineer, Prompt Engineer) scored as weighted SSOC blends

License & Credits

MIT License. Adaptable for other countries via ISCO-08 crosswalks.

Made by Kirill So with Claude (Anthropic) & Codex (OpenAI). Data from MOM, O*NET, Felten et al. (2021), Pizzinelli et al. (2023), Anthropic Economic Index, Eloundou et al. (2023/2024), ILO, and Stanford DEL.

← Home

Methodology Appendix Data