Every source has a publisher, author, publication date, method, geography, limitation, and verification date. Most importantly, this library distinguishes evidence already used in the published seed model from newer work that informs interpretation but has not changed a score.
Current evidence status
Occupation tasks
31.0 (August 2026) Current O*NET production release, integrated across all 620 public occupations.
National wages
May 2025 Current OEWS release: 602 publishable medians across 620 profiles.
Published scores
seed-v0.4-2026-05 6 source families are declared in the versioned seed model.
Research library
17 verified sources Canonical links checked through 9 September 2026.
Evidence roles
Two lanes, one visible boundary
A source can be excellent research without being part of the current scoring formula.
01 / Published seed score6 source families
These source IDs appear in seed-v0.4-2026-05 and in published occupation records. Card-level counts below show exactly how many of the 620 records cite each source.
02 / Interpretation and calibration queue11 newer sources
These studies update how we discuss exposure, realized employment, task reorganization, global differences, and retraining. Their evidence role is explicit: none currently calculates a public score.
Newest evidence
Interpretation and calibration queue
The queue now includes 2030 macroeconomic scenarios, current official exposure methodology, observed payroll and hiring signals, cross-country infrastructure differences, cross-vendor usage evidence, and experimental retraining research.
Interpretation / calibration queueThe Anthropic Institute
Tracked for interpretation and the next documented calibration review; not used to calculate current scores.
Authors
Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter, Peter McCrory
Coverage
United States, 2026-2030 scenario horizon
Method
Task-based macroeconomic model mapping AI capability, adoption, productivity, automation, new-task creation, and worker reallocation assumptions to GDP, wages, labor share, employment, and unemployment through 2030.
Limit
The three paths are illustrative scenarios without probabilities, use only two broad occupation groups, omit policy responses and several macroeconomic feedbacks, and do not predict outcomes for individual occupations.
Provenance
Publisher canonical URL
Verified
Interpretation / calibration queueAnthropic Economic Research
Tracked for interpretation and the next documented calibration review; not used to calculate current scores.
Authors
David Roodman, Maxim Massenkoff
Coverage
United States and Europe
Method
Meta-analysis of 56 randomized US studies, supplemented with experimental evidence from Europe.
Limit
Historical training programs may not reproduce at AI-transition scale, and average effects do not guarantee a good fit for a specific worker or local market.
Provenance
Publisher canonical URL
Verified
Interpretation / calibration queueStanford Digital Economy Lab
Tracked for interpretation and the next documented calibration review; not used to calculate current scores.
Authors
Erik Brynjolfsson, Bharat Chandar, Ruyu Chen
Coverage
United States
Method
High-frequency ADP administrative payroll records covering millions of US workers through June 2026, compared across age and occupational AI-exposure groups.
Limit
The employment patterns are early descriptive indicators, not causal estimates of AI-driven displacement.
Version note
Revised August 2026 edition; replaces the original August 2025 working paper for current interpretation.
Provenance
Publisher canonical URL
Verified
Interpretation / calibration queueOpenAI Economic Research
Tracked for interpretation and the next documented calibration review; not used to calculate current scores.
Authors
OpenAI
Coverage
United States ChatGPT workplace sample
Method
Privacy-preserving analysis of more than 800,000 work-related messages from US ChatGPT users, classifying generic and occupation-specific task crossover.
Limit
Observed product use is an early signal of task reorganization, not a representative employment forecast.
Provenance
Publisher canonical URL
Verified
Interpretation / calibration queueOpenAI Economic Research
Tracked for interpretation and the next documented calibration review; not used to calculate current scores.
Authors
OECD
Coverage
OECD occupational frameworks
Method
Maps AI capabilities across nine cognitive, social, and physical domains to occupational requirements and constructs an updateable AI Capability Gap index.
Limit
Capability proximity measures potential exposure over a five-to-ten-year horizon; it does not measure adoption, employment effects, or displacement timing.
Version note
Modernizes occupational AI-exposure measurement; retained alongside the broader OECD Employment Outlook 2023 because their scopes differ.
Tracked for interpretation and the next documented calibration review; not used to calculate current scores.
Authors
Paweł Gmyrek, Mariana Viollaz, Hernan Winkler
Coverage
135 countries covering around two-thirds of global employment
Method
Cross-country occupational exposure analysis combining task composition, digital connectivity, income group, sector, and automation-versus-augmentation potential.
Limit
Country-level infrastructure and occupational structure clarify uneven exposure but do not establish realized job loss or local hiring effects.
Version note
Extends the ILO 2025 exposure index with digital-infrastructure and cross-country task differences; it does not replace the underlying index.
Tracked for interpretation and the next documented calibration review; not used to calculate current scores.
Authors
Maxim Massenkoff, Peter McCrory
Coverage
United States
Method
Combines theoretical model capability with observed Claude usage, weighting work-related automation more heavily, then compares exposure with BLS projections and CPS outcomes.
Limit
Claude usage is not representative of all AI use, and the reported youth-hiring signal is suggestive rather than causal.
Version note
Applies Anthropic's observed-use primitives to labor-market outcomes; it complements rather than replaces the January 2026 primitives report.
Provenance
Publisher canonical URL
Verified
Interpretation / calibration queueAnthropic Economic Index
Tracked for interpretation and the next documented calibration review; not used to calculate current scores.
Authors
Kiran Tomlinson, Sonia Jaffe, Will Wang, Scott Counts, Siddharth Suri
Coverage
Global Bing Copilot usage mapped to US occupations
Method
Maps 200,000 anonymized, privacy-scrubbed Bing Copilot conversations to work activities and O*NET occupations, measuring breadth and successful assistance.
Limit
Copilot users are not a representative labor-force sample, and AI applicability must not be interpreted as job elimination.
Provenance
Publisher canonical URL
Verified
Current model inputs
Evidence used in the published seed score
These are the only research and official-data sources currently represented in role-level score provenance. A citation count indicates use, not causal proof or equal weighting.
Cited by 40 of 620 published occupation score records.
Authors
World Economic Forum
Coverage
55 economies and 22 industry clusters
Method
Survey of more than 1,000 global employers representing over 14 million workers, combined with partner labor-market datasets for the 2025–2030 outlook.
Limit
Survey-based projections reflect employer expectations, not realized displacement.
Provenance
Publisher canonical URL
Verified
Used in published seed scoreInternational Labour Organization
Cited by 414 of 620 published occupation score records.
Authors
Paweł Gmyrek, Janine Berg, Karol Kamiński, Filip Konopczyński, Agnieszka Ładna, Balint Nafradi, Konrad Rosłaniec, Marek Troszyński
Coverage
Global
Method
Task-level assessment of nearly 30,000 tasks at six-digit occupation detail, combining human expert input and AI predictions into four exposure gradients.
Limit
Exposure estimates do not directly predict layoffs or wage outcomes.
Cited by 620 of 620 published occupation score records.
Authors
US Bureau of Labor Statistics
Coverage
United States
Method
Official employer survey estimates of employment and wage distributions by occupation; displacement.ai imports the May 2025 national release by SOC code.
Limit
Wage and employment figures are historical labor-market measures, not AI-specific forecasts.
Cited by 620 of 620 published occupation score records.
Authors
O*NET Resource Center, US Department of Labor
Coverage
United States
Method
Official O*NET-SOC occupation taxonomy and task, skill, knowledge, and work-activity records maintained through survey and occupational analyst processes.
Limit
Task descriptions require modeling to translate into AI exposure and displacement pressure.
Provenance
Publisher canonical URL
Verified
What the 2026 evidence changes—and what it does not
Economic scenarios: Anthropic's new task-based model links capability, adoption, automation, productivity, and worker reallocation assumptions to conditional 2030 outcomes. Its three paths are scenarios without probabilities, not occupation forecasts.
Observed employment: Stanford payroll evidence finds no economy-wide displacement signal, while documenting a widening employment gap for workers aged 22–25 in exposed occupations. The study is descriptive, not causal.
Exposure measurement: OECD's nine-domain capability-gap method is more transparent and updateable than a static automation-risk label, but it still measures capability rather than adoption or job loss.
Observed AI use: Anthropic, Microsoft, and OpenAI usage studies show what people attempt with specific products. Product populations are not representative labor-force samples.
Global variation: ILO and World Bank research shows that connectivity, task composition, institutions, and income level change the balance between automation and augmentation.
Worker transitions: Randomized retraining evidence finds positive but modest average effects. Recommendations should validate local demand, credentials, cost, and worker circumstances.
Source policy
Use canonical publisher pages and record authors, publication date, access date, geography, method, and limitation.
Treat capability, exposure, observed product use, employer expectations, and realized employment as different evidence classes.
Do not convert exposure or AI applicability into an employment-loss claim without labor-market outcome evidence.
Keep US national evidence visibly separate from local wages, licensing, hiring demand, and non-US labor markets.