How we evaluate AI claims
A transparent method for separating measured change from technical exposure and long-range prediction.
Evidence hierarchy
- Administrative employment and earnings data
- Controlled experiments and field studies
- Observed AI usage and task data
- Job postings, business surveys, and industry evidence
- Task-exposure models
- Executive claims and forecasts
Lower-ranked sources can still be useful. They simply cannot prove what stronger evidence is required to establish.
Our three evidence labels
Observed means directly measured by a credible source. Strong inference means multiple evidence streams support the conclusion without measuring it directly. Speculative describes a plausible scenario with substantial uncertainty.
Tasks before titles
An occupation is a bundle of tasks, relationships, responsibilities, physical actions, and legal obligations. AI can automate several tasks without eliminating the occupation. It can also let fewer workers handle more output, changing hiring even when the job title survives.
Corrections and updates
Material corrections identify what changed, why, when, and whether the conclusion changed. Permanent field guides display a research cutoff or last-reviewed date.