--- title: "Labor market impacts of AI: A new measure and early evidence" aliases: - "Anthropic Observed Exposure" - "Massenkoff & McCrory 2026" - "Anthropic Labor Market Note" tags: - us - descriptive - exposure-measure - age-gradient - hiring - anthropic source: "/research/2026-03-anthropic-labor-market-impacts.pdf" extract: "/research/extracts/2026-03-anthropic-labor-market-impacts.md" authors: "Maxim Massenkoff; Peter McCrory" publisher: "Anthropic" date: "2026-03-05" --- # Labor market impacts of AI: A new measure and early evidence > [!info] Source > PDF (human): [2026-03-anthropic-labor-market-impacts.pdf](/research/2026-03-anthropic-labor-market-impacts.pdf) · Raw extract (machine): [2026-03-anthropic-labor-market-impacts.md](/research/extracts/2026-03-anthropic-labor-market-impacts.md) > Anthropic research note — 5 March 2026 — Empirical research note introducing the **Observed Exposure** measure, 17 pages > URL: https://www.anthropic.com/research/labor-market-impacts > **Data:** Anthropic Economic Index (CC BY 4.0) + O*NET + US Current Population Survey + Eloundou et al. β ## TL;DR Anthropic introduces a new measure of AI displacement risk — **Observed Exposure** — that combines theoretical LLM capability with real-world Anthropic usage data, weighting automated and work-related uses more heavily. Actual coverage remains a fraction of theoretical. Occupations with higher observed exposure are projected by BLS to grow less through 2034. **No systematic increase in unemployment for highly exposed workers since late 2022**, though there is suggestive evidence that hiring of younger workers (22–25) has slowed in exposed occupations (≈14% drop in job-finding rate post-ChatGPT, just barely significant). ## Key findings - **Theoretical vs actual capability gap is large.** β=1 (fully feasible) tasks are 68% of observed Claude usage, β=0 are just 3%. But Claude currently covers only ~33% of all tasks in Computer & Math (where β covers 94%); Office & Admin covers ~90% by β. Adoption is far behind technical capability. - **Top 10 most exposed occupations** under Observed Exposure include Computer Programmers (75% coverage), Customer Service Representatives, and Data Entry Keyers (67%). - **30% of workers have zero coverage** — Cooks, Motorcycle Mechanics, Lifeguards, Bartenders, Dishwashers, Dressing Room Attendants. - **Validation against BLS:** every 10pp increase in coverage → 0.6pp drop in BLS 2024–2034 employment growth projection (weighted regression). The Eloundou-only measure shows no such correlation — Observed Exposure adds signal. - **High-exposure workers are different:** in the top quartile vs zero-exposure group, +16pp female, +11pp white, ~2× as likely to be Asian, **earn 47% more**, much higher education (graduate degrees: 17.4% vs 4.5%, ~4× difference). - **Unemployment:** no statistically significant divergence between top-quartile and zero-exposure groups since ChatGPT release. CI permits detecting differential ~1pp increases in unemployment. - **Younger workers show suggestive divergence.** Job-finding rate (CPS panel) for 22–25 year olds in high-exposure occupations falls ~14% post-ChatGPT (barely statistically significant). No such effect for 25+. Possible alternative interpretations: not hired but staying in existing job, taking different jobs, or returning to school. - **Detection thresholds for hypothetical scenarios:** if the top 10% of coverage were laid off, top-quartile unemployment would jump from 3% → 43%, aggregate from 4% → 13%. A "Great Recession for white-collar workers" (top-quartile unemployment doubling 3% → 6%) would be detectable. ## Methodology in brief Combines three sources: (1) O*NET task taxonomy (~800 occupations, 18,400 tasks), (2) Anthropic Economic Index — task-level Claude usage, (3) Eloundou et al. β scores. Counts a task as covered if it is theoretically capable AND has sufficient work-related Claude usage; weights automated implementations fully, augmentative at half. Aggregates to occupations weighted by time-fraction. Compared to BLS projections; analysed against CPS unemployment & hiring data with difference-in-differences vs zero-exposure baseline. ## Implications for AdaptAI ### Calculator (`/calculator`) - The 10 most-exposed occupations list should be cross-checked against `jobTitlesWeighted` for: Computer Programmers (-5 currently — paper supports raising to a less protective number for entry-level roles, but keep architects/seniors low), Customer Service Representatives (entry-level admin roles already weighted high), Data Entry Keyers (very high weight justified). - The 14% drop in young-worker hiring is a strong signal for **age-sensitive insights** — calculator could surface a specific risk message for users under 25 in high-exposure occupations. ### Big Picture (`/big-picture`) - `bigPictureData.industryExposure.aiExposure` — Observed Exposure validates higher exposure for Admin & Customer Support, Tech & IT, Finance. - `bigPictureData.unemploymentByAge` — paper finds null for 25+, suggestive negative for 22–25. Refresh framing copy on `/big-picture` to highlight this age-cohort divergence; it is consistent with the existing higher youth unemployment trajectory in the dataset. - `bigPictureData.graduateEmployment` — supports the projection that graduate hiring is falling faster than total employment. ### Industry Insights - `/industries/tech` — Computer Programmers at top of exposure list at 75% coverage. Refresh the prose with this stat. - `/industries/admin-customer-support` and `/industries/admin-support` — Customer Service Reps and Data Entry Keyers in top 10 most exposed. Strong supporting citations available. ### How to Adapt - `/adapt/upskilling` — the methodology shows that **observed adoption is a fraction of theoretical**, meaning workers have time to upskill. Useful counterweight to fatalist narratives. - `/adapt/career-transitions` — explicit advice for younger workers most affected by hiring slowdown. - `/adapt/human-advantage` — interactions are 29% task iteration, 28% directive, 24% learning, 11% feedback, 6% validation. Validation/judgement remains scarce in the data — supporting human-advantage framing. ## Related notes - [[CBP UK Evidence Review]] — uses the same Anthropic Economic Index but applied to UK occupations. - [[GLA Working Paper 103]] — alternative ILO task-based methodology applied to London-specific data. - [[Same Storm, Different Boats]] — independent Swedish-register corroboration of the young-worker hiring slowdown. ## Caveats & limitations - US data only; UK applicability needs cross-walking SOC2020 ↔ O*NET. - Anthropic Economic Index is Claude-only; under-counts ChatGPT/Gemini/Copilot usage. - "Observed" is not "displaced" — high adoption may indicate augmentation, not replacement. - The 14% young-worker hiring decline is "barely statistically significant"; the paper itself flags this. - Confidence intervals only permit detecting ≈1pp differential unemployment changes — smaller real effects could be hidden. - 30% of workers have zero coverage simply because their tasks didn't reach the minimum sample threshold, not because they are inherently safe.