Evidence monitor / AI & work
What the evidence
actually measures.
Studies of AI can measure different populations, outcomes and mechanisms. This record keeps those distinctions attached to the findings.
Last source review: September 9, 2026 · Selected studies, not an exhaustive or systematic review
Current assessment. Evidence warrants close attention to early-career hiring and changing tasks. The reviewed studies do not establish widespread, economy-wide displacement caused by AI.
| Study & version | Design & population | Finding | Limit on the claim |
|---|---|---|---|
| Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence Revised August 12, 2026; data through June 2026 | Employment association Young workers in U.S. ADP payroll data; observed through June 2026 | Employment diverges by AI exposure for younger workers; hiring is the main channel. | A relative gap is not a national count of jobs destroyed or a causal estimate of AI displacement. |
| Large Language Models, Small Labor Market Effects 2025 working paper; linked Danish survey and administrative data | Linked survey and administrative study Danish workers in selected exposed occupations | Task change and adoption coexist with small early effects on recorded earnings and hours. | Different institutions and outcomes from the U.S.; average effects can mask subgroup changes. |
| Do Job Postings Show Early Labor-Market Effects of AI? May 14, 2026 | Vacancy analysis U.S. online job postings, classified by occupational AI exposure | The posting evidence does not identify a distinct new national decline attributable to AI. | Pre-existing trends and other demand shocks complicate attribution; ads are not hires. |
| Job postings show early signs of AI automation impact September 1, 2026 | Vacancy analysis U.S. postings; automation-focused task exposure, with a Texas application | More exposed positions show weaker posting demand in the model. | Exposure definition differs from broader AI indices. Posting estimates do not measure actual separations. |
| Generative AI and Jobs: A Refined Global Index of Occupational Exposure May 20, 2025 | Task exposure index Occupations across the global workforce | Task overlap suggests substantial potential for job transformation. | Capability overlap is not measured adoption, employment loss or a forecast of individual risk. |
| Generative AI at Work 2025 published article | Workplace rollout study · peer reviewed Customer-support agents in one studied workplace | AI assistance improves measured productivity, especially for less experienced workers. | Does not establish economy-wide effects on staffing, pay or the distribution of productivity gains. |
Capability → adoption → outcomes
A tool completing a task establishes a capability. A worker or employer using it establishes adoption. Changes in staffing, hiring, earnings or hours establish outcomes. Attributing those outcomes to AI requires a research design that addresses other explanations.
Each step creates a new question. Treating them as equivalent produces both exaggerated fears and premature reassurance.
Forecast accountability
A forecast record needs the original dated statement, the population, horizon, measurable outcome and conditions for evaluating it. An executive’s prediction belongs in that record as a prediction. It should not enter the observed employment count.
Opening’s reporting agenda includes a forecast ledger. It is not yet a completed scored dataset; ambiguous forecasts will remain explicitly unscored. See the reporting agenda.