Why AI Hasn’t Increased Unemployment, According to Anthropic

What It Covers

Anthropic's head of economics Peter McCrory examines why AI has not increased US unemployment despite 20% of firms adopting AI in at least one business function, presenting an 18-month research framework explaining AI's labor-augmenting rather than displacing effects, with caveats about future agentic capabilities.

Key Questions Answered

• •Labor augmentation over displacement: AI functions as a skill-biased, labor-augmenting technology because no occupation in the Department of Labor's O*NET taxonomy has all its associated tasks fully handled by Claude. Essential non-automated tasks — interpersonal coordination, physical presence, complex judgment — constrain displacement while amplifying returns to human expertise working alongside AI.
• •Expertise compounds AI output: Anthropic's research across 81,000 Claude users found that sophisticated user inputs correlate directly with complex, higher-quality Claude outputs. After six months of use, workers increasingly treat Claude as a reasoning partner rather than a tool, and this behavior pattern produces measurably more successful interactions — suggesting domain expertise grows more valuable, not less.
• •Jobs are not fixed task bundles: Historical precedent and current Anthropic data show that technology reshapes task combinations within roles rather than simply eliminating roles. Workers report a primary productivity gain as expanded scope — doing more, more proficiently. This task rebundling dynamic means AI raises the marginal product of labor even while automating specific sub-tasks.
• •Watch hiring rates, not layoffs: The most detectable early signal of AI labor impact appears in hiring slowdowns, not unemployment spikes. Anthropic's data shows weakened hiring rates for young workers in high-AI-exposure roles over the past year — consistent with Stanford Digital Economy Lab findings — while overall unemployment at 4.2% remains at levels the Fed considers full employment.
• •Agentic coding preserves expertise premium: Analysis of Claude Code usage over seven months shows that as AI handles more implementation work, human planning and delegation skills become more valuable, not less. Workers with stronger domain expertise recover from Claude errors more consistently and succeed on complex tasks more often, meaning agentic AI raises returns to judgment over raw coding ability.