Divya van Mahajan

Why AI Hasn’t Increased Unemployment, According to Anthropic

ai-daily-brief-podcast
Listen to episode →

Why AI Hasn’t Increased Unemployment — According to Anthropic

Overview

This episode of the AI Daily Brief (recorded around July 24, 2026) covers a range of AI industry news before diving into a detailed discussion of a post by Peter McCrory, Head of Economics at Anthropic, titled “Why Hasn’t AI Increased Unemployment?” The host synthesizes McCrory’s analysis, contextualizes it with broader economic data, and explores responses from technologists, investors, and researchers. The central argument is that AI, so far, functions as a skill-biased, labor-augmenting technology rather than a labor-displacing one — and the host examines why that is, what might change it, and what it means for how leaders should think about deploying AI.

Note: No direct YouTube URL was provided for this episode.

Prerequisites

  • Basic understanding of macroeconomics and labor market indicators (unemployment rate, employment-to-population ratio, job openings ratio)
  • Familiarity with the concept of general-purpose technologies (GPTs) and historical analogies (e.g., electrification, the internet)
  • Awareness of major AI labs and their flagship models (Anthropic/Claude, OpenAI/GPT series)
  • Understanding of terms like “task automation,” “labor augmentation,” and “skill-biased technological change”
  • Familiarity with Jevons’ Paradox (briefly referenced) and the concept of economic singularities
  • General awareness of the current AI product landscape (agents, agentic coding, model routing)

Main Points

Headlines: Model Routing Is Becoming a First-Class Feature

  • Stripe is reportedly in talks to acquire OpenRouter for ~$10 billion — a dramatic markup from its $1.3 billion valuation just two months prior.
  • The strategic rationale: Stripe gains the “metering and billing layer for inference” and a developer funnel; OpenRouter benefits from enterprise distribution.
  • Simultaneously, Cursor launched its own model router, allowing engineers to optimize for intelligence, cost, or balance — claiming frontier-level performance at 60% cost reduction in intelligence mode.
  • Multiple companies (Meta, Ramp, Vercel) are also building routing products, signaling that model routing has shifted from a niche tool to a core infrastructure feature.

Headlines: Voice Mode Expansions at Anthropic and OpenAI

  • Anthropic expanded voice mode to Opus and Sonnet models (previously limited to Haiku), allowing users to handle complex tasks through voice without sacrificing model quality.
  • Voice now integrates with apps like Gmail, Slack, and Notion; foreign language support moved out of beta.
  • OpenAI launched voice in its desktop app via a new real-time model called GPT-Live, capable of background task execution during natural conversation.

Headlines: Amazon Retreating from Frontier AI; Microsoft Doubling Down

  • Amazon’s AGI division (founded 2023, led briefly by David Luan) is experiencing significant layoffs and reportedly shutting down its dedicated spin-off lab; the Nova model family has seen no new release since December.
  • By contrast, Microsoft published reinforcement-learning results from its MAI model family, showing that fine-tuning a small coding model (CodeOne Flash) within product-specific harnesses (GitHub Copilot, Excel) can deliver near-frontier performance at a fraction of the cost.
  • Microsoft reported an 84% reduction in cost replacing OpenAI’s image model with MAI Image 2.5 in PowerPoint; Satya Nadella articulated a strategy of routing lower-complexity tasks to cheaper in-house models while reserving frontier models for frontier needs.

Headlines: Policy, Security, and Geopolitical AI Tensions

  • Bipartisan legislation introduced in the U.S. House (“AI Kill Switch Bill”) would require AI companies to maintain shutdown capability and grant DHS authority to issue shutdown commands.
  • Secretary of State Rubio pushed back diplomatically, urging allies not to pursue digital sovereignty programs that would reduce dependence on U.S. platforms.
  • A joint U.S.-U.K. evaluation found Kimi K3 significantly behind U.S. frontier models on cybersecurity benchmarks (32.2% vs. 76.2% average for U.S. models on ExploitBench); concern about Chinese AI capabilities was characterized by some officials as overstated.
  • Anthropic and OpenAI were identified as the primary industry advocates for restricting open-source models with strong cyber capabilities, drawing criticism from NVIDIA CEO Jensen Huang, who argued model diversity is safer than a concentrated duopoly.

Main Episode: The McCrory Framework — Why Hasn’t AI Caused Unemployment?

  • U.S. labor market baseline is healthy: June 2026 unemployment at 4.2% (consistent with Fed’s definition of full employment); prime-age employment-to-population ratio near multi-decade highs; initial unemployment claims stably low for four years.
  • AI adoption is large enough that macroeconomic effects should be detectable: 20% of firms use AI in at least one business function; quality-adjusted AI output grew over 2,000% per year in both 2024 and 2025; labor productivity growth rose from 1.6% (pre-pandemic) to 2.0% per year (2022–2026).
  • No evidence of material job displacement in high-exposure roles: Workers in roles where Claude automates a large share of tasks do not show elevated unemployment relative to lower-exposure workers; hiring of young workers in high-exposure roles has weakened, but this is likely attributable to broader macroeconomic volatility rather than AI specifically.

Main Episode: Why AI Is a Skill-Biased, Labor-Augmenting Technology

  • No job in the O*NET taxonomy has all its tasks fully automated by Claude — even heavily automated roles retain non-automatable elements requiring interpersonal coordination, physical presence, or complex judgment.
  • Sophisticated inputs correlate with complex outputs: Anthropic data shows that when Claude produces high-value work (e.g., economic models), a human with domain expertise is directing the process — AI amplifies rather than replaces that expertise.
  • Persistent returns to expertise in agentic coding: Analysis of Claude Code usage over seven months showed that people with more domain expertise succeed more often and recover from Claude errors more consistently; “the return to straightforward coding ability may have fallen, but agentic coding has so far increased the value of other complementary skills.”
  • Jobs are not fixed task bundles: Historically, automation reshapes job content rather than simply eliminating jobs — AI may automate some tasks, reinforce others, and create entirely new task bundles, increasing the marginal product of labor on net.
  • User psychology shifts with experience: Among 81,000 Claude users, those with more usage experience decrease their expectations of job loss and grow more optimistic about AI’s impact on pay and job security, even as they increase their expectations of what Claude can do.

Main Episode: What Could Change — Risks and Caveats

  • As the “jagged frontier” of model capability smooths out, the skill-biased, labor-augmenting character of AI may weaken; autonomous agents handling complex, long-horizon tasks could alter the current equilibrium.
  • Recursive self-improvement / AI-automated innovation is the most disruptive scenario: standard economic models suggest this could produce singularities (infinite growth in finite time) — though “weak links” (tasks that are essential and hard to automate) may constrain this.
  • The most likely near-term labor market signal is not unemployment but changes in hiring patterns: fewer junior roles, smaller teams, slower backfilling, higher per-employee expectations — overall unemployment may remain stable even as workforce composition shifts.
  • McCrory’s personal forecast: unemployment will not be “noticeably higher” one year from now due to AI.

Host Commentary: Narrative Framing Matters

  • The dominant narrative around AI’s economic impact has a self-reinforcing effect on executive behavior: if the story is “cut headcount,” leaders will cut headcount; if the story is “do more, move faster, expand scope,” leaders will pursue expansion.
  • The host explicitly advocates for the augmentation narrative as both more empirically grounded (given current evidence) and more conducive to positive economic outcomes.

Key Concepts

  • Model Routing: The practice of automatically directing AI queries to different models based on task requirements, cost, or performance targets — increasingly treated as core infrastructure.
  • Token Scarcity / Token Budget: A shift in enterprise AI strategy from maximizing model capability on all tasks to tightly controlling inference spending per task.
  • Skill-Biased Technological Change: A pattern in which a new technology raises the productivity and wages of skilled workers more than unskilled workers, increasing inequality in returns to labor.
  • Labor-Augmenting Technology: Technology that increases the effective productivity of human workers rather than substituting for them entirely.
  • Jagged Frontier: The uneven capability profile of current AI models — highly capable in some areas, surprisingly weak in others — requiring human expertise to navigate gaps.
  • O*NET Taxonomy: The U.S. Department of Labor’s catalog of occupations and their associated tasks, used as a framework for measuring AI exposure across job categories.
  • Agentic Coding (Claude Code): Use of AI systems to autonomously execute multi-step coding tasks under human oversight, shifting the human role toward planning and delegation.
  • Recursive Self-Improvement (RSI): The hypothetical process by which AI systems automate the development of more capable AI systems, potentially producing nonlinear or singular economic growth dynamics.
  • Weak Links (economic concept): Essential tasks or processes that are difficult to automate and therefore constrain overall productivity growth even under widespread automation, as described by Agion, Jones, and Jones.
  • Hill-Climbing (Microsoft’s term): An iterative fine-tuning methodology in which a model is trained within a specific product harness using real user feedback to progressively improve task-specific performance.
  • Frontier Tuning: Microsoft’s service allowing customers to fine-tune MAI base models for custom enterprise applications.
  • ExploitBench / The Last Ones: Cybersecurity evaluation benchmarks used to assess an AI model’s ability to autonomously conduct network intrusion attacks.
  • Jevons’ Paradox: The historical observation that increases in the efficiency of resource use often lead to greater total consumption of that resource, referenced briefly as relevant to AI productivity dynamics.

Summary

The central argument of this episode, drawn primarily from Peter McCrory’s analysis at Anthropic, is that AI has not materially increased unemployment in the United States despite reaching a scale large enough that macroeconomic effects should be detectable. The U.S. labor market remains near full employment, and workers in roles heavily exposed to AI automation show no elevated unemployment relative to less-exposed workers. McCrory’s explanation is that current AI exhibits the hallmarks of a skill-biased, labor-augmenting technology: no job’s full task bundle is yet automated, sophisticated AI outputs consistently require sophisticated human direction, and returns to domain expertise have remained persistent even in agentic contexts. Jobs are not static task bundles, and AI is reshaping job content and expanding scope rather than simply eliminating roles. Caveats remain: young workers’ hiring may be softening, the most disruptive effects may appear first in hiring rates and team composition rather than in unemployment statistics, and recursive self-improvement could eventually alter the current equilibrium. The host concludes that the framing adopted by leaders and the public — augmentation versus headcount reduction — is not merely descriptive but actively shapes how AI gets deployed, making the propagation of an augmentation narrative both empirically justified by current evidence and consequentially important for the kind of economic future that results.