The Entry-Level Signal: What Anthropic's Observed-Exposure Data Reveals About Early-Career Hiring

See also (wiki): ai-talent-workforce-planning · chro-ai-workflows

Executive Summary

• Anthropic economists Massenkoff and McCrory built a novel “observed exposure” measure — combining O*NET task data with actual Claude usage patterns — and tracked it against CPS employment data post-ChatGPT. The finding is more precise than prior theoretical exposure indices: only 33% of Computer & Math tasks are actually being covered by AI despite 94% theoretical feasibility.
• Aggregate unemployment has not increased for workers in high-exposure occupations since late 2022. The economy-wide signal that AI is eliminating jobs at scale is not yet present in the data.
• The exception is early-career workers. Job-finding rates for 22-25 year olds entering AI-exposed occupations fell by approximately half a percentage point post-ChatGPT — a 14% decline in hiring rate. No similar effect exists for workers over 25.
• This matches the BCG Henderson Institute finding (April 9, 2026) that 61% of the most automation-vulnerable roles are entry-level. Two independent methodologies reach the same place.
• The workforce planning implication is specific: companies that have historically depended on hiring two junior analysts per senior role, or relied on entry-level roles as the first rung of a career ladder, face a structural pipeline problem that will widen as observed AI coverage expands.

What Makes This Measurement Different

The core problem with most AI labor market research is the gap between what AI can do and what AI is actually doing. Prior exposure indices (including the widely cited Eloundou et al. 2023 GPT-4 task assessment) estimate theoretical capability — how much of a given job could an LLM perform if deployed. These measures consistently find that high-skill, high-wage workers in knowledge-economy roles face the highest theoretical exposure.

Massenkoff and McCrory add an observed dimension: they cross-reference O*NET task descriptions against actual Claude usage patterns (the Anthropic Economic Index) to estimate which tasks AI is actually covering in real workflows. The gap is stark.

Computer & Mathematical occupations: 94% theoretical exposure, 33% observed coverage. Roughly two-thirds of the tasks LLMs are theoretically capable of handling are not being routinely handled by AI in practice. This is consistent with a world where AI adoption is real but uneven — where the model can do the work but workflows have not been redesigned to let it.

30% of workers have zero observed AI coverage in their occupation. The theoretical exposure maps a ceiling; the observed coverage maps where the economy currently sits.

Source credibility: MEDIUM-HIGH. Massenkoff and McCrory are Anthropic economists — the employer has a direct commercial interest in AI being perceived as impactful, which creates incentive to both overstate observed coverage (to demonstrate AI relevance) and understate labor displacement (to reduce regulatory and public-relations pressure). Both biases are possible simultaneously. The methodological contribution — using actual usage data rather than theoretical task matching — is genuinely valuable and addresses a real gap in the literature. Cross-reference against: UMD/LinkUp behavioral job-posting data (April 2026), Fed FEDS regression (Liu & Webber, March 27, 2026), and BCG Henderson task-level taxonomy (April 9, 2026). The demographic findings (47% wage premium for exposed workers, entry-level hiring decline) converge with independent sources and are the most credible outputs of this paper. TIER 1 (March 5, 2026).

Who Is Actually Exposed

The demographic profile of workers in high-observed-exposure occupations (pre-ChatGPT baseline, Aug–Oct 2022) is not the intuitive picture:

AI is not concentrating in low-wage, low-skill occupations. The most AI-exposed workers are the most credentialed, the highest-paid, and — at 16pp above baseline female share — more likely to be women. Computer programmers (75% coverage) and data entry workers (67%) sit at the high end of the observed-coverage distribution; customer service representatives also appear at elevated levels.

This matters for two reasons. First, organizations running AI impact assessments that only audit low-wage clerical roles are looking in the wrong place. The highest observed automation coverage is in knowledge-work roles. Second, the equity dimension cuts differently than most AI-displacement narratives assume — the workers with the most to lose from observed AI coverage are the workers with the most structural advantages in the labor market.

The Aggregate Finding: No Signal Yet

The paper’s most reassuring finding is also its most methodologically constrained one: there is no statistically significant increase in unemployment for workers in highly exposed occupations since late 2022. The differential unemployment change — exposed workers vs. unexposed workers — is “small and insignificant.”

The methodological constraint matters. The analysis detects differential changes in exposed occupations relative to others. It would not detect parallel unemployment increases that hit all sectors simultaneously. And the detection threshold is approximately 1 percentage point — effects smaller than that are in the noise. A scenario where exposed workers face, say, a 0.7pp higher unemployment rate than unexposed workers would be missed.

What the paper establishes: AI has not yet caused a measurable, sector-specific unemployment surge. What it does not establish: that no displacement is occurring within jobs, that hours and wages are unaffected, or that aggregate employment-level effects don’t exist below detection thresholds.

The Early-Career Signal

The most actionable finding for workforce planning is the 22–25 age cohort. Job-finding rates for young workers entering AI-exposed occupations fell by approximately half a percentage point post-ChatGPT — representing a 14% decline in hiring rate. No similar effect exists for workers over 25.

The paper characterizes this result as “barely statistically significant” — a caution worth honoring. But it converges with three independent data sources that reach the same directional conclusion:

• BCG Henderson Institute (April 9, 2026): 61% of roles most vulnerable to automation replacement are entry-level or junior positions. The learn-as-you-go career pathway is the most structurally disrupted.
• AEI’s Brent Orrell (April 14, 2026): Observed dramatic, sustained decline in entry-level coding job postings over the past six to seven years; demand for senior engineers and architects held steady.
• UMD/LinkUp (April 2026): Entry-level job postings are at an 8-year high across all categories — but the composition is shifting away from roles that AI can fill toward roles that require human judgment at entry.

The 14% decline in hiring rate for 22-25 year olds into exposed fields is a leading indicator, not a crisis. Young workers who don’t get hired into those roles may shift fields, continue education, or take unexposed positions. But the pipeline implication for organizations is concrete: if you have historically recruited entry-level talent into roles with high observed AI coverage, those roles are contracting — and the candidates you expected to hire are being absorbed elsewhere.

Key Data Points

What This Means for Your Organization

The workforce planning takeaway from this paper is not “AI isn’t affecting jobs” — that would misread the aggregate finding. The correct read is: AI has not yet produced visible aggregate unemployment, but it is already reshaping which jobs exist at the entry level, and it is doing so in the high-knowledge-work roles most mid-market organizations depend on for talent pipelines.

Three practical implications.

First, audit your entry-level hiring by observed AI coverage, not theoretical exposure. The 33% actual coverage on Computer & Math tasks (vs. 94% theoretical) means that blanket assumptions about which roles are at risk are unreliable. A role that looks automation-exposed on paper may have the majority of its actual daily tasks untouched by AI today. Role-level task analysis using behavioral data — not theoretical assessments — produces plannable conclusions.

Second, the demographic profile of exposed workers changes the equity math. High-exposure workers are more educated, higher-paid, and disproportionately female. An AI rollout that reduces headcount in high-observed-coverage roles is not a low-skill displacement event. It is a knowledge-worker event with specific demographic concentration. HR and legal teams building AI impact assessments need updated assumptions about who bears the workforce burden.

Third, the entry-level pipeline problem is early-stage and addressable. A 14% decline in hiring rates for 22-25 year olds in exposed fields is a signal, not a collapse. Organizations that restructure their entry-level intake now — building junior roles around judgment, client-facing work, and AI-oversight tasks rather than data entry and routine analysis — will have access to mid-market talent that their competitors are inadvertently screening out. That is a concrete competitive opportunity.

If this raised specific questions about which of your roles fall into which observed-coverage buckets, or how to reclassify entry-level intake around AI-adjacent work, I’d welcome the conversation — brandon@brandonsneider.com.

Sources

Brandon Sneider | brandon@brandonsneider.com April 2026