In March 2026, Anthropic published research on AI’s labour market impacts using their own Claude usage data. They measured occupational exposure and employment effects. The data is damning. The framing is cope.
“The Impact of AI on Labour Markets” — Anthropic Research, March 5, 2026. Uses Claude API usage data mapped to O*NET occupational categories to measure exposure and employment effects.
Anthropic’s headline finding: “no systematic increase in unemployment” in AI-exposed occupations. Sounds reassuring. But buried in the same paper:
Younger worker hiring has slowed in exposed occupations
New entrants to AI-exposed fields are finding fewer positions. The jobs aren’t being eliminated yet — they’re just not being created. Attrition without replacement. The displacement happens through non-hiring, not firing.
Exposure is concentrated in high-skill, high-pay occupations
Unlike previous automation waves that hit manual labour, AI exposure is concentrated in knowledge work. Programmers, analysts, customer service — the middle class backbone.
“No systematic unemployment” ≠ “no displacement”
Unemployment is a lagging indicator. When a company stops hiring juniors because Claude handles their work, unemployment stats don’t move — but labour demand has been permanently reduced. Anthropic measured the wrong metric and declared victory.
| Dimension | OpenAI (Signals) | Anthropic (Labour Research) |
|---|---|---|
| Data Source | ChatGPT aggregate usage telemetry | Claude API usage + BLS employment data |
| Time Period | July 2024 – March 2026 (21 months) | 2024 – early 2026 |
| Key Metric | 46% of work messages are “doing” | 75% programmer exposure |
| Job Activity Mapping | 165 O*NET IWA codes | O*NET occupational categories |
| Displacement Finding | Implicit (framed as adoption) | Explicit but minimised |
| Cope Strategy | Rename displacement as adoption | Measure lagging indicator, declare no problem |
| Cope Score | 72 • HEAVY | 58 • MODERATE |
Anthropic scores lower on cope than OpenAI because they at least acknowledged displacement mechanisms exist. They published actual labour market analysis, not just usage metrics. But concluding “no systematic unemployment” while documenting hiring slowdowns in exposed occupations is textbook cope — just better-written cope.