Methodology
How This Site Works
Every number on jobsdata.ai traces back to a cited source. This page explains how we collect, weight, combine, and present evidence across every section of the site.
TL;DR
jobsdata.ai is a live meta-analysis of AI’s impact on jobs, not a forecast. We track 470+ individually cited sources across 16 prediction graphs, a 5-dimensional occupation risk framework, and real-time automation signals. Peer-reviewed research (Tier 1) counts 8× more than blog posts (Tier 4). Newer studies get up to 1.5× more weight. The headline number on every graph is a weighted average that updates automatically when new research drops. The site covers displacement, wages, adoption, and corporate behavior across the US economy.
Overview
TL;DR
This site synthesizes published research into a structured dashboard. It is not a model, not a forecast, and not a single team’s opinion. When sources disagree, the disagreement is shown, not hidden.
jobsdata.ai tracks how predictions about AI’s impact on employment, wages, and work itself evolve as new evidence emerges. The site organizes research across three layers:
- Prediction graphs (16 graphs across displacement, wages, and adoption) — each plotting every published estimate we can find, weighted by evidence quality and recency.
- Occupation risk framework (342 occupations scored across 5 dimensions) — going beyond simple AI exposure to account for adoption speed, worker adaptability, demand elasticity, and whether AI replaces or enhances.
- Automation signals (150+ developer tools tracked by industry) — package downloads as a leading indicator of where AI automation is heading, like construction permits for software.
Each layer has its own methodology detailed below. They share a common evidence tier system and a commitment to showing uncertainty honestly.
Prediction Graphs
TL;DR
16 graphs covering displacement (8), wages (4), and adoption (4). Each graph plots every published estimate as a data point, color-coded by evidence tier. The headline number is a weighted average across all sources. You can toggle tiers on and off to see how the number shifts.
The prediction graphs are the core of the site. Each tracks a specific metric (e.g., “Projected US Job Displacement from AI by 2030”) across all published estimates we can find.
Evidence Tiers
TL;DR
Four tiers of source quality. Tier 1 (peer-reviewed, government data) gets 8× the weight of Tier 4 (blogs, social media). The dashboard works without Tier 4. You can filter by tier on every graph.
Peer-reviewed studies (RCTs, journal articles, NBER working papers), government statistics (BLS, Census, OECD), SEC filings with legal liability for accuracy.
Think tanks (Brookings, McKinsey, RAND), international organizations (IMF, ILO, World Bank), working papers (arXiv, SSRN, IZA), industry research (Gartner, Forrester).
Major outlets (WSJ, FT, Fortune, CNBC). Cited when they report original data or on-the-record disclosures. Powers the news feed.
Twitter/X threads, Substack, blogs, podcasts. Used sparingly for ground-level signals. The dashboard works without them.
How We Calculate the Headline Number
TL;DR
The large number on every prediction tile is a weighted average. Four factors determine each data point’s influence: evidence tier, recency, sample size, and whether it’s a direct or proxy measurement.
The headline number is not a median, not a simple average, and not cherry-picked. It’s a weighted mean where better evidence counts more:
mean = ∑(value × weight) / ∑(weight)
Tier 1 = 4×, Tier 2 = 2×, Tier 3 = 1×, Tier 4 = 0.5×. This means a single peer-reviewed study has the same influence as eight blog posts. When you toggle tiers on the graph, the weighted average recalculates instantly with only the selected tiers.
Proxy Metrics
TL;DR
Many studies measure something related to but not identical to a graph’s unit. Job posting declines are not the same as job losses. We convert proxies using empirically grounded factors, widen confidence bands, and apply a 0.5× weight discount so direct measurements always dominate.
Plotting a 12% job posting decline on a displacement graph where direct measurements cluster around 3–6% creates a false outlier. The proxy methodology prevents this by converting indirect measurements to the target unit before plotting.
| Proxy Metric | Target | Factor | Range | Basis |
|---|---|---|---|---|
| Job posting decline (%) | % jobs displaced | 0.35 | 0.20–0.50 | Postings drop 2–3× faster than actual layoffs (Cajner et al. 2020) |
| Task automation potential (%) | % jobs displaced | 0.30 | 0.15–0.50 | Only ~30% of automatable tasks lead to job restructuring within 5 years (OECD 2023) |
| Relative posting change (%) | % jobs displaced | 0.30 | 0.15–0.45 | Relative comparisons overstate net displacement via composition effects |
| Productivity gain (%) | % wage change | 0.40 | 0.20–0.60 | Historical pass-through ~40% medium-term (Stansbury/Summers 2020) |
| Revenue automation (%) | % jobs displaced | 0.25 | 0.10–0.40 | Revenue automation often precedes reallocation, not reduction |
5-Dimensional Displacement Risk
TL;DR
AI exposure alone doesn’t predict displacement. We score 342 occupations across five research-backed dimensions: technical exposure (30%), adoption speed (20%), worker adaptability (15%), demand elasticity (15%), and AI complementarity (20%). The first two drive displacement up. The last three absorb it.
Most AI-labor analyses stop at “exposure” — what share of tasks can AI do? That’s necessary but insufficient. An occupation can be highly exposed to AI and still see employment growth if demand is elastic, workers can adapt, and AI is deployed as a complement rather than a substitute. The 5-dimensional framework captures all five factors.
Question: How many of this occupation’s tasks can current or near-term AI actually perform?
Method: GPT-scored analysis of 342 BLS occupations, validated against six independent academic exposure indices (Yale Budget Lab meta-analysis, Pearson r = 0.878).
Sources: Eloundou et al. (2023); Felten, Raj & Seamans (2023); Yale Budget Lab (2026); Acemoglu (2024).
Key insight: Acemoglu’s bounding argument shows that ~20% of tasks are AI-exposed, ~23% can be profitably automated, with ~27% cost savings per task — yielding only 0.53–0.66% TFP gain over a decade. Exposure alone dramatically overstates likely displacement.
Automation Signals
TL;DR
We track 150+ Python/JavaScript packages as leading indicators of AI automation. Think of it like construction permits: before AI replaces tasks in an industry, developers download the tools to build those systems. The signal is in relative growth rates, not absolute numbers.
Before a building goes up, construction permits spike in that neighborhood. Before AI automates tasks in an industry, developers start downloading the tools to build those automation systems. We track both Python (PyPI) and JavaScript (npm) package downloads as a leading indicator of where AI automation is heading.
J-Curve Framework
TL;DR
Measured productivity lags real AI impact because the enormous complementary investments required (reorganization, retraining, workflow redesign) are expensed, not capitalized. This creates a J-shaped curve: productivity looks worse before it looks better. Electricity took ~40 years. Computers took ~25.
The J-Curve framework from Brynjolfsson, Rock & Syverson (2021, AEJ: Macroeconomics) explains why AI’s productivity impact hasn’t shown up in the macro data yet. General-purpose technologies require massive complementary investments that national accounts treat as expenses, not capital.
Demand Elasticity & the Jevons Paradox
TL;DR
When AI makes work cheaper, total demand for that work often expands. ATMs made branches cheaper to run, so banks opened more branches. Teller employment grew for 30 years. Occupations with elastic demand may see more hiring from AI, not less.
The Jevons Paradox, first observed in 1865 with coal, applies directly to AI-labor economics: when efficiency makes a product or service cheaper, consumption often increases enough to offset the efficiency gain.
Historical Context: GPT Cycles
TL;DR
AI follows the same arc as every prior general-purpose technology: emergence, diffusion, displacement, reorganization, new equilibrium. Each GPT has taken 25–40 years from emergence to productivity surge. The diffusion timeline may be compressing because software replicates faster than physical infrastructure.
- Emergence — New technology introduced, narrow applications. (AI: 2012–2022)
- Diffusion — Rapid adoption, experimentation across sectors. (AI: 2023–present)
- Displacement — Old occupations and skills lose value. (AI: emerging, uneven)
- Reorganization — Firms restructure workflows, new job categories emerge.
- New equilibrium — Productivity surges, labor reallocation stabilizes.
Research Pipeline
TL;DR
An automated pipeline queries 11 academic databases weekly, deduplicates results, and ranks papers by tier, relevance, and citation velocity. 35 leading AI-labor researchers are monitored continuously. Human review gates every ingestion.
New research is surfaced automatically but never ingested automatically. The pipeline finds candidates; a human decides what enters the dataset.
Research digest runs weekly (Mondays). News ticker refreshes hourly. Prediction data updates when new evidence materially changes an estimate.
Known Limitations
TL;DR
This is a map of published research, not an oracle. The weighted average masks disagreement. The 5-dimensional weights are author-chosen, not econometrically estimated. Coverage is English-language only. These limitations are features of transparency, not bugs to be hidden.
Definitional heterogeneity. Sources within a single prediction may define the metric differently. “Displacement” can mean roles eliminated, tasks automated, or jobs restructured depending on the study. The weighted average treats these as equivalent, which may overstate or understate the true range.
Point estimates from wide ranges. The headline number can mask significant disagreement. When the source range is wide (e.g., 0–12%), the range itself is often more informative than the midpoint. We show the full range on every prediction page.
Trend arrows reflect source mix, not reality. A downward arrow may mean “newer sources estimate lower,” not that the real-world metric is declining. Adding or removing a source can shift the trend.
5-dimensional weights are author-chosen. The 30/20/15/15/20 weighting is not econometrically estimated. Demand elasticity classifications are qualitative. CFO complementarity data covers only 8 of 22 occupation groups. The framework’s value is showing that single-dimension exposure scores omit most of what determines actual displacement, not in the precise composite number.
Within-group variation. The occupation framework operates at the SOC major group level (22 groups), which masks significant variation. “Computer and mathematical” includes both data entry clerks and machine learning researchers.
Binary replacement risks not captured. The task-composition model assumes gradual, task-by-task automation. It does not capture cases where a single breakthrough automates the dominant task wholesale (e.g., autonomous vehicles for trucking). If that capability crosses the threshold, the complementarity buffer disappears.
Coverage gaps. English-language sources only. Keyword matching can miss or misclassify papers. Citation data lags publication, so new work may be underranked. Signal data (package downloads) is inflated by CI/CD pipelines and does not equal production use.
Structural tendencies, not individual predictions. These scores describe group-level patterns. A person’s actual risk depends on their specific role, employer, geography, skills, and adaptability — not just their occupation group’s average.
Know a study we should include? Suggest a source and we’ll review it for inclusion.
Update schedule: Research digest runs weekly (Mondays) across all 11 sources. News ticker refreshes hourly. Prediction data updates when new evidence materially changes an estimate. Signal data (package downloads) updates monthly.