V5 is the live structural score: a deterministic 3-layer model with latent-source posterior exposure, task-mode adjustment, human bottleneck, and market resilience. Prior V4.2 baselines and V4.3 shadow artifacts remain published for auditability. V5 adds transition-adjusted risk, realized-risk layers, and calibrated uncertainty diagnostics.
Live Site | Methodology | Calculator | Data
- 562 occupations scored across 9 major groups
- 88 synthetic roles (product manager, data scientist, delivery rider, startup founder...)
- 27% face high+ AI risk (152 occupations)
- SGD 56B Est. annual wage pool under pressure
- 50% average AI task overlap across all occupations
- 4-source exposure ensemble: Felten AIOE + Anthropic observed usage + Eloundou GPT exposure + ILO 2025
Three-layer deterministic scoring — no LLM in the pipeline:
- Exposure — reliability-weighted 4-source ensemble: AIOE (2021), Anthropic Economic Index (2026), Eloundou GPTs-are-GPTs (2024), ILO Refined Index (2025), blended using latent-source posterior in V5.
- Human Bottleneck — Pizzinelli theta from O*NET Work Context (judgment, presence, coordination)
- Market Resilience — MOM employment/wage trends + SOL/JiD demand signals + Anthropic calibration
net_risk = exposure × (1 − bottleneck) × market_modifier
Published as risk bands (Very Low through Very High) with visible confidence, retained V4.2 baselines, and task-adjusted bootstrap/stability diagnostics.
- BLS cross-country: live convergent cross-check against US BLS projections
- Cluster-level: directional check against Singapore labour-monitor clusters
- Release pipeline: validated end to end with published artifacts, checksums, claims matrix, and shadow-model governance outputs
- Methodology page: aiworkindex.pages.dev/methodology
git clone https://github.com/kirso/aiworkindex
cd aiworkindex
bun install
bun run build:release-data # Refresh all release datasets and metadata
bun run scripts/score.ts # Score all 562 occupations
bun run validate # Run release and data-contract validation
bun run dev # Start dev server
bun run build # Build 672 prerendered pages| Source | What | Year |
|---|---|---|
| MOM Singapore | 562 SSOC occupations, wages, employment | 2024-2025 |
| Felten AIOE | AI exposure per SOC (academic index) | 2021 |
| Anthropic Economic Index | Observed AI usage (HuggingFace, CC-BY) | Jan 2026 |
| Eloundou et al. | GPT-4 task-level exposure (Science, 2024) | 2024 |
| ILO Refined Index | ISCO-08 exposure (52K expert data points) | May 2025 |
| O*NET | Work Context, Job Zones, Task Statements | 2020 |
| MOM SOL 2026 | Shortage Occupation List | Nov 2025 |
| MOM Jobs in Demand | In-demand occupation flags | Dec 2025 |
| US BLS | Employment projections 2024-2034 (convergent cross-check) | Aug 2025 |
Each occupation page shows:
- Education level (O*NET Job Zones → Singapore labels)
- Progressive Wage Model coverage (57 occupations in 9 PWM sectors)
- Licensed profession flag (53 strict + 23 partial)
- Foreign worker dependency (73 very high + 33 high + 45 moderate)
- SkillsFuture career conversion eligibility (154 occupations)
- Industry footprint + worker profile from official Singapore labour tables
- Transition infrastructure from Jobs Transformation Maps, CareersFinder, WSQ, and SkillsFuture / WSG programmes
- Public registry: aiworkindex.pages.dev/research
- Machine-readable artifact:
static/data/research-library.json - Live methodology references are now generated from the same canonical research registry used by reports and release governance.
- SvelteKit 5 + Svelte 5 runes (static site, adapter-static)
- Tailwind CSS v4 + shadcn-svelte (Bits UI)
- D3.js for visualization layout
- TypeScript scoring pipeline (Bun runtime)
- Satori + Resvg for OG image generation
- Deployed on Cloudflare Workers
- Exposure data age: AIOE is from 2021 (pre-GPT-4). Ensemble with newer sources mitigates but doesn't eliminate.
- Employment granularity: detailed Singapore occupation counts are not publicly released.
estimated_sg_employment_thousandsis a labeled sub-major allocation, and wage-pool analysis uses a separate labeled BLS-weighted proxy. - Market modifier weights: 0.6/0.4 momentum/scarcity split is calibrated, not empirically derived.
- Cluster-level labour data: Same vacancy/hiring data for all occupations in each of 3 clusters.
- Synthetic role weights: Expert-assigned SSOC blends, not validated against job posting data.
MIT