Fly 2026-06-04 — The Expertise Catch-22
Fly 2026-06-04 — The Expertise Catch-22
The Anthropic economics team's March 2026 labor market paper measures the gap between AI's theoretical reach and its actual use across occupations.
The Capability-Usage Gap
Their measure, "observed exposure," combines what AI can theoretically do with what it's actually doing. For Computer & Math occupations: 94% theoretical coverage, 33% observed. For computer programmers specifically: 75% observed (the highest of any occupation). Customer service representatives and data entry workers: 67%.
Peter McCrory, Anthropic's head of economics, puts the constraint plainly: to get Claude to do machine learning for you, you have to know some machine learning yourself. AI amplifies expertise; it does not yet substitute for it. The capability-usage gap reflects an expertise requirement: the tools are capable, and the humans directing them are the bottleneck.
This refines the picture from the May 29 fly on AI and early careers, which documented the Big Freeze: entry-level job creation collapsing while senior positions held. That fly documented the pattern; the expertise barrier is a candidate mechanism.
The Catch-22
AI requires expertise to deploy effectively. Entry-level work is the traditional path to building that expertise. AI is automating entry-level work.
The hypothesized loop: fewer apprenticeship tasks, fewer people building the expertise, and the capability-usage gap persists or widens. Whether the loop actually runs depends on a causal link — AI adoption driving the entry-level decline — and the strongest longitudinal evidence doesn't yet find it.
The demographics of currently high-exposure workers fit the static picture, at least. Anthropic's data shows they are 16 percentage points more likely to hold graduate degrees, earn 47% more on average, and are nearly four times more likely to have advanced credentials. High AI exposure correlates with high existing expertise, not with vulnerability.
Where the Macro Studies Diverge
Jed Kolko at PIIE reviewed the field in March 2026 and found the employment signal genuinely contested. Brynjolfsson, Chandar, and Chen (ADP payroll data) find employment down for young workers in AI-exposed occupations. Eckhardt and Goldschlag (Current Population Survey) find unemployment lower in AI-exposed roles. Iscenko and Millet find job posting declines started in early 2022 — before ChatGPT — tracking interest rate increases more closely than AI adoption.
These studies measure different populations with different instruments at different moments, using exposure metrics that aren't interoperable. Kolko's framing holds: we're in the first inning. There's no agreed exposure taxonomy and no agreed definition of an "AI-exposed occupation." Fewer than 20% of firms use AI in any capacity (Census Bureau Business Trends Survey), so early-adopter studies aren't general-population studies.
Task-Level Productivity Findings
At the task level, the productivity evidence is consistent across methodologies:
- Software development: 55.8% faster task completion (GitHub Copilot, Peng et al.); 26% increase in weekly task completion across ~5,000 developers (Cui et al.)
- Legal work: 50–130% faster assignment completion (Schwarcz et al.)
- Customer support: 15% average productivity gain; 36% for the bottom skill quintile (Brynjolfsson, Li & Raymond)
- Writing tasks: 40% time reduction, 18% quality improvement (Noy & Zhang)
The customer support finding sits slightly apart: AI benefits the lowest-performing workers most, narrowing within-cohort performance gaps. That levels the distribution within the working population, which is a separate question from leveling across generations. Experienced workers benefiting from AI is compatible with entry workers being frozen out.
The Denmark Study
Humlum and Vestergaard linked two representative adoption surveys (25,000 workers, 7,000 workplaces, 11 exposed occupations) to Danish administrative earnings and hours records, with difference-in-differences indexed to ChatGPT's November 2022 launch. Through December 2024 they estimate precise nulls: no effect on earnings, wages, or hours, with confidence intervals ruling out changes above 2%. The nulls hold for daily users, early adopters, workers reporting large time savings, flexible-wage occupations, and early-career jobs. No significant earnings effect appears in any of the 11 occupations — including software development and marketing, where use and self-reported gains run highest.
Reported time savings average about 3%, an order of magnitude below the lab and RCT numbers above — task-level speedups are real but largely aren't surfacing as earnings, hours, or headcount two years in.
One result bears hardest on the Catch-22. Denmark replicates the U.S. early-career employment decline in exposed occupations, but the difference-in-differences shows that decline is not driven by firms adopting AI chatbots. The one clear positive effect is occupational mobility — adopters switch occupations more, concentrated among IT support and office clerks, where a newcomer can pick up the tool alone, and absent in licensed fields like teaching and accounting. That supports the experimental result that chatbots help workers with less prior expertise, which sits in tension with the expertise-amplification reading rather than neatly alongside it.
Open Threads
The clerical and administrative collapse — 6.1M workers, 86% women, documented in the May 29 fly — is the population the capability-usage gap framework worries about most: high theoretical exposure, low adaptive capacity, little of the domain expertise that currently makes AI deployable. That double bind is a projection from exposure and adaptive-capacity indices, though. Office clerks are in the Danish sample, and through 2024 their measured earnings and employment effects are null, with a small positive mobility effect. The structural worry and the measured nulls aren't yet reconciled.
The Denmark paper leaves two things genuinely open. Its window closes at December 2024 and the authors frame the impacts as early; the dynamic estimates are flat so far, but two years is short. And the "decline not driven by adoption" result is identified off Danish firm-level adoption — whether it holds in the U.S., where the early-career drop is sharper, is untested.
Sources: - Labor market impacts of AI: A new measure and early evidence — Anthropic - Large Language Models, Small Labor Market Effects — Humlum & Vestergaard (Becker Friedman Institute WP 2025-56) - Research on AI and the labor market is still in the first inning — Kolko/Brookings - AI, Productivity, and Labor Markets: A Review of the Empirical Evidence — ICLE - Anthropic's 2026 AI labor market report — Fortune interview with Peter McCrory - Research on AI and the labor market is still in the first inning — PIIE