Research & methodology · model task-curve-v1
How we know what we say — and how we guess what we can't know.
What the scores mean
The headline number on every profile is the Task Transformation Index (TTI): the weighted share of an occupation's task content that is performed, or substantially assisted, by machines in a given year and scenario, from 0 to 100.
It is not the probability that the job disappears. Studies that report “X% of jobs at risk” measure technical automatability, and their results vary from about 47% to 9–14% depending on whether they account for the variety of tasks within an occupation. We follow the task-based approach for that reason, and report employment as a separate, qualitative outlook.
History shows automation of tasks can coexist with growing employment — bank tellers after the ATM are the classic case — or end an occupation outright, as with telephone operators .
The task model
- Decompose. Each occupation is broken into 3–10 weighted tasks, structured from O*NET task statements where available, and classified into seven kinds following the routine/non-routine framework of Autor, Levy & Murnane .
- Record 2026. Each task gets an observed state for 2026: human-led, assisted, automated, retired (with year) or emerging (with expected year). Before 2026, intensity is reconstructed from the documented year technology began assisting the task.
- Capability curves. After 2026, each task kind follows a logistic capability curve: the year machines reach half their ceiling (midpoint), the spread (width), and the ceiling. These are editorial parameters informed by AI-exposure research .
- Adoption lag. Capability is not adoption. Each occupation has five barrier factors (0–1) that delay adoption: lag = 3 + 14·regulation + (9 or 2)·environment + 6·social + 5·(1 − economics) years, scaled by scenario.
- Accountability cap. Judgement tasks are capped by (1 − 0.75 × accountability) and interpersonal tasks by (1 − 0.55 × social acceptance).
- Never below observed. A forecast never falls below the 2026 observed state of a task.
| Task kind | Midpoint | Width (yrs) | Ceiling |
|---|---|---|---|
| Routine cognitive | 2024 | 6 | 0.95 |
| Routine physical | 2030 | 10 | 0.92 |
| Non-routine analytic | 2038 | 9 | 0.85 |
| Creative production | 2042 | 12 | 0.7 |
| Non-routine physical (unstructured settings) | 2060 | 14 | 0.78 |
| Interpersonal & care | 2068 | 18 | 0.45 |
| Judgement & accountability | 2078 | 20 | 0.35 |
Employment outlook combines the change in transformation since 2026 with each occupation's underlying demand trend (−1 to +1: e.g. ageing populations raise demand for care; net-zero lowers demand for coal mining) and the share of newly emerging tasks. Where demand is elastic, productivity can increase employment. The outlook is a direction — Growing, Stable/transforming, Reinvented, Contracting, Sharply contracting — never a headcount.
Resilience score (0–100) = 80% retained human task content in 2050 (central) + 20% demand trend.
Three scenarios
Conservative
Technology develops more slowly, with significant economic, regulatory and physical constraints.
midpoints +14 yrs · ceilings ×0.82 · adoption lag ×1.45 · new tasks ×1.5
Central
Technology progresses along plausible adoption pathways based on current research and observed developments.
midpoints +0 yrs · ceilings ×1 · adoption lag ×1 · new tasks ×1
Accelerated
Major breakthroughs lead to faster automation, robotics adoption and AI integration.
midpoints -9 yrs · ceilings ×1.1 · adoption lag ×0.6 · new tasks ×0.6
Every occupation is assessed against the same factors: technical feasibility (capability curves), economic viability, adoption barriers, regulatory constraints, physical environment, human accountability, social acceptance, infrastructure requirements (within adoption lag) and skills transition (emerging tasks). You can set your own assumptions in Build your own future.
Confidence classification
- High confidenceWell-documented by multiple authoritative sources.
- Moderate confidenceSupported by evidence, with some uncertainty about scale or timing.
- Low confidencePlausible, based on trends; timing and magnitude are uncertain.
- SpeculativeAn exploratory scenario beyond the reach of reliable forecasting.
Confidence falls with the horizon and is capped by evidence depth: 2026 is High for detailed profiles; 2030 is at most Moderate; 2050 is Low; anything beyond 2050 is Speculative. A prediction for 2126 never carries the same reliability as one for 2030. Future events on the timeline are always labelled Scenario — not a fact.
Evidence standards
- Only real, identifiable publications. No invented citations, statistics or events. Every source has a unique ID, publisher, author, publication date, URL, evidence type, access date and confidence class — see the source register (178 sources, 547 citations).
- Priority to primary and authoritative sources: ILO, OECD, IMF, World Bank, ONS, BLS, O*NET, IFR, UNESCO, WHO, IEA, peer-reviewed research, and original historical documents.
- Each source records what it says (summary) and how not to read it (caveat). Opinion and industry estimates are labelled as such and never treated as established fact.
- Where sources disagree we show the disagreement — see “Will AI create widespread unemployment?” on the human impact page.
- Automated verification. A verifier checks every URL (HTTP status) and resolves every DOI through Crossref, confirming the registered title. Status is stored per source and displayed; publishers that block automated checks are flagged rather than hidden.
- Where evidence is unavailable we say so: profiles without an occupation-specific narrative are marked Core; salary forecasts are not published; historical imagery is withheld until licensing is confirmed.
Taxonomy & coverage
Occupations are organised as family → subfamily → occupation, and mapped to ISCO-08 unit groups, O*NET-SOC codes and, progressively, UK SOC 2020 . The initial dataset holds 147 occupations in 24 families: 19 detailed profiles (occupation-specific milestone narratives) and 128 core profiles.
All content lives as reviewed, typed data, is validated by automated tests (every reference must resolve; scenarios can never be High confidence) and is seeded into PostgreSQL. The interface contains no occupation content, so new occupations need no design work.
Limitations
- Barrier factors and capability parameters are editorial judgements informed by research, not estimated statistically. They are published so they can be challenged.
- The dataset does not yet hold employment counts, so cross-occupation averages are unweighted.
- Occupation definitions are mostly UK/US-centric; task content differs across countries and income levels .
- Generative AI is too recent for its long-run labour market effects to be measured .
- Beyond 2050 everything is speculative. The value is in comparing occupations under the same assumptions, not in the absolute numbers.
Roadmap to comprehensive coverage
- Classification backbone: import the full ISCO-08 structure (436 unit groups) and the O*NET-SOC taxonomy (about 900+ occupations) as skeleton profiles marked Incomplete.
- Task import: map O*NET task statements and work activities to the seven task kinds to generate draft task lists for review.
- Employment weights: attach BLS OEWS and ONS employment counts so averages can be employment-weighted.
- History: derive 1920/1930 occupational counts from IPUMS census microdata to anchor 1926 snapshots quantitatively.
- AI-assisted drafting with human review: the modular AI provider drafts profiles only from supplied, verified sources; drafts enter a review queue and are never published automatically.
- Internationalisation: country-specific factors (regulation, income level) and translated profiles.