← narwal.one/Second Brain
SecondBrain
Ask the Brain
Index/Sourceupdated Tue Jul 28 2026 08:00:00 GMT+0800 (Philippine Standard Time)

Why AI Hasnt Increased Unemployment According to Anthropic (AI Daily Brief)

sourceai-and-jobslabor-economicsanthropicaugmentationmodel-routingunemployment

Why AI Hasn't Increased Unemployment, According to Anthropic (AI Daily Brief)

The AI Daily Brief main episode (~2026-07-24, captured 2026-07-28): Nathaniel Whittemore reads and reacts to Peter McCrory's (head of economics at Anthropic; views-his-own disclaimer noted) X essay "Why hasn't AI increased unemployment?" — the fullest first-party statement yet of the Skill-Biased Labor Augmentation thesis. Headlines: Stripe in talks to buy OpenRouter for $10B, model routing goes first-class, Microsoft's in-house model strategy pays off, Amazon shuts its AGI lab.

The McCrory essay — core argument

Claim: the US labor market is stable and close to maximum employment; "AI has caused no material increase in the unemployment rate to date" — even for workers highly exposed to current patterns of AI automation.

Setup — this absence is informative, not premature: the AI sector is big enough that effects should be visible. 20% of firms use AI in ≥1 business function (40% in the information sector, 5.5% of GDP); quality-adjusted AI output grew >2,000%/yr in 2024 and 2025; labor productivity growth ran 2%/yr 2022–2026 vs 1.6% pre-pandemic (he reads AI into that). So "no unemployment signal yet" is a real finding, not noise.

On displacement evidence: mixed, and he's unconvinced. Anthropic's labor-impact report finds no worsening unemployment for workers in roles whose tasks Claude is used to automate (updated with recent BLS data, result holds). There IS "suggestive evidence that hiring rates for young workers in highly AI-exposed roles have weakened" — consistent with the Stanford Digital Economy Lab's Canaries in the Coal Mine paper (Erik Brynjolfsson's lab) — but he cautions: AI arrived during "the largest non-recessionary labor market slowdown on record" (post-pandemic unwinding, rate tightening, trade-war uncertainty), and a low-hire/low-fire market hits early-career entrants hardest for reasons other than AI.

Why no unemployment? AI is skill-biased and labor-augmenting:

  • Capabilities are "stubbornly jagged" — expert oversight fills the pockets of the jagged frontier and recovers when models falter
  • No O*NET occupation has all its tasks systematically handled by Claude; the non-automated remainder both constrains the productivity lift and amplifies returns to labor
  • Sophisticated user inputs and complex Claude outputs are highly correlated — complex work happens under expert direction
  • Jobs are not fixed bundles of tasks — task automation redraws role boundaries and re-bundles work while raising the marginal product of labor (see Tasks to Responsibilities Shift)
  • Claude Code data (7 months): value of tasks rose, but persistent returns to human expertise — people with more domain expertise succeed more often and recover more consistently from Claude's errors. "The return to straightforward coding ability may have fallen, but agentic coding has so far increased the value of other complementary skills."
  • Exposure is uneven: technical writers, data entry, customer support, computer programmers are projected by BLS to grow less through 2034 — augmentation-in-aggregate coexists with pockets of real displacement risk

The caveats he keeps: exposed-role workers do express greater fear of job loss (see FOBO (Fear of Becoming Obsolete)); the skill-bias could dissolve as the jagged frontier smooths; and recursive self-improvement is the wild card — citing Aghion, Jones & Jones: "economic growth may be constrained not by what we do well, but rather by what is essential and yet hard to improve" — essential non-automated "weak links" can keep the labor share elevated even under rapid, incomplete automation.

His prediction: "I don't expect unemployment to be noticeably higher a year from now, at least not because of AI."

Reactions Whittemore surfaces

  • Andy Hall (Stanford): the real political backlash was predicted for unemployment +2pp; "we're still not there, and this piece helps explain why."
  • Trace Cohen: the impact may show up in hiring, not layoffs — fewer junior roles, smaller teams, slower backfilling, higher expectations per employee. Whittemore agrees team size is the metric to watch (smaller, more nimble teams with expanding responsibilities).
  • Julie Fredrickson / Robert Scoble: notable simply as Anthropic saying something positive about AI and work — cf. the ad-campaign misfire in AI Optimism vs AI Pessimism (AI Daily Brief).

Whittemore's closing thesis — narratives are self-fulfilling: if the story executives absorb is "use AI to cut staff in half," investors expect exactly that; if it's "doing more faster and moving into lateral domains," leaders deploy AI expansively. The augmentation narrative isn't just description — it shapes adoption behavior. (The mirror-image mechanism of the AI Layoff Trap, where the replacement narrative is the self-fulfilling one.)

Headlines segment

  • Stripe in talks to acquire OpenRouter for ~$10B (WSJ) — vs $1.3B valuation two months earlier; framed as buying "the metering and billing layer for inference." Token-scarcity-era logic: the best router wins big. See Model Routing, Token Scarcity.
  • Cursor Router launched (intelligence/cost/balanced modes; claims Fable-level performance at −60% cost); Meta, Ramp, and Vercel also shipped routers. Matthew Berman: "model routing is a first-class feature now."
  • Anthropic voice mode extended from Haiku to Opus + Sonnet models, with connector support (Gmail/Slack/Notion); OpenAI shipped desktop voice on the GPT live realtime model.
  • Amazon shuts its AGI lab (the SF spin-off; Nova model family stalled since December, leadership departures); spokesperson denies exiting model training. Andrew Curran: "Amazon is giving up on Nova, would be my guess."
  • Microsoft's MAI "hill-climbing machine" — MAI Code 1 Flash post-trained in the GitHub Copilot harness beat GPT-5.4 Mini and Haiku 4.5 on accept rate (+10%) with 10% lower token usage; Excel-harness training pushed SWE-bench Verified 72%→86%; MAI Image 2.5 replaces GPT Image 2 as PowerPoint/Bing default (−84% cost per Suleyman). Satya Nadella: optimize "the cost to outcome frontier" — route to MAI for saturated capabilities, frontier models for frontier needs.
  • SpaceX AI expanding earth-side data centers — Texas site at Colossus scale (~1GW), Pentagon compute talks. See SpaceX.
  • AI Kill Switch Bill (Reps. Ted Lieu + Nathaniel Moran): AI companies must maintain shutdown/throttle/suspend capability; DHS gets shutdown-command authority. Rubio's diplomatic cable simultaneously instructs diplomats to deny the US would arbitrarily cut allies off. See Kill Switch.
  • CAISI/UK AISI joint eval of Kimi K3 — big lag vs US frontier on cyber benchmarks (32.2% vs 76.2% avg on ExploitBench); Lutnick: "the Kimi panic needs to stop." See Moonshot AI.
  • Anthropic regulatory-capture discourse — The Information: Anthropic and OpenAI are essentially the only industry voices advocating a distillation/open-model crackdown; Jensen Huang counter-argues a single-model monoculture is "much, much more vulnerable" than a myriad of open models.
  • Meta's AI-optimism ad campaign — "we're betting on people"; Whittemore: generic but welcome (pairs with the Anthropic ad critique in AI Optimism vs AI Pessimism (AI Daily Brief)).

Connects to