Data & methods
Measure each stage on its own terms.
The observatory separates adoption, recurring use, assisted working time, reported counterfactual savings, and realized outcomes. Each construct has its own denominator and evidentiary status.
Measurement
Five constructs anchor the observatory.
- Work adoption
- Share of employed adults aged 18–64 reporting generative AI use for their job.
- Used for work last week
- Share of the employed population reporting work use in the prior week.
- Used every workday
- Share of the employed population reporting GenAI use on every workday.
- Work hours assisted
- Share of total work hours actively using GenAI, reconstructed from reported use duration and days worked. Nonusers contribute zero active-use hours.
- Reported time savings
- Respondents' counterfactual estimate of the additional time the same work would have required without GenAI. Nonusers contribute zero reported saved hours.
Interpretation boundary
Reported time savings are survey-based counterfactual estimates. Labor productivity, output, GDP, and employer value added require direct outcome evidence.
Analysis
Technical detail stays available on demand.
Longitudinal diagnostics
Subgroup results are evaluated wave by wave over the complete common A/H/S window, currently Q4 2024 through Q2 2026. Cross-sectional Pearson and Spearman correlations describe alignment between constructs. Reported R² values are descriptive fits across aggregate groups. Worker-level probabilities and causal estimates require different designs.
Rank stability uses Spearman correlations between subgroup rankings for every pair of audited quarters. The publication reports the median, endpoint, minimum, and maximum pairwise correlations across the full window.
Occupation-composition analysis
The composition layer estimates an industry value under national occupation-specific GenAI rates combined with that industry's occupation mix. This is a standardization counterfactual. Causal decomposition requires separate identification.
- Adoption counterfactual
- Uses CPS worker-share occupation weights within each industry.
- Assisted-hours counterfactual
- Uses actual-main-job-hour occupation weights within each industry.
- Reported-savings counterfactual
- Uses the same actual-main-job-hour weighting basis as the work-hour outcome.
- Residual
- Observed industry value minus its occupation-composition counterfactual. The code names this an
occupation_.adjusted_ industry_ context_ residual
The CPS composition foundation uses official Q2 2025 and Q2 2026 inputs. Release 1 reports occupation-adjusted industry-context residuals as derived descriptive diagnostics. Organizational quality, efficiency, productivity, and causal firm effects sit outside the supported interpretation. Design-based confidence intervals for the custom pooled CPS composition vectors remain unsupported.
Evidence and release boundary
RPS/FRED supplies the registered aggregate measurements. Longitudinal diagnostics and composition counterfactuals form derived descriptive evidence. Outcome and causal inference require additional evidence.
The authorized release pipeline retrieves the published aggregate RPS source into a private candidate workspace, validates the registered 131-series inventory, and builds deterministic publication artifacts. Release 1 uses DATA_MODE=derived_only and excludes private source-input bytes from the public bundle.
Source identities, crosswalk versions, validators, and interpretive guardrails are versioned with the code. A source-definition, rights, or release-state change blocks publication until the candidate passes the project's release controls.