Anthropic Labor Market Impacts of AI: New Exposure Measure and Early Evidence

TL;DR

Anthropic released a new "observed exposure" measure that blends theoretical LLM capability with actual usage data to assess AI displacement risk, showing that AI has not yet raised unemployment rates but may be reducing hiring for workers aged 22‑25 in the most exposed jobs.

New "observed exposure" metric

Key point: Observed exposure quantifies the share of an occupation’s tasks that are both theoretically feasible for an LLM and actually used in automated, work‑related contexts on Anthropic’s platforms.

The metric builds on three data sources: the O*NET task database, Anthropic’s Economic Index usage logs, and task‑level exposure scores (β) from Eloundou et al. (2023).
Tasks receive full weight if used via fully automated APIs and half weight if used augmentatively.
Occupation‑level exposure is a time‑weighted average of task‑level coverage.

"Observed exposure" is meant to quantify: of those tasks that LLMs could theoretically speed up, which are actually seeing automated usage in professional settings?

How observed exposure differs from theoretical capability

Key point: AI usage is far below the theoretical ceiling; only a fraction of tasks rated β = 1 or β = 0.5 are actually automated.

Figure 1 (Anthropic) shows that 68 % of Claude usage falls on tasks with β = 1, while tasks with β = 0 account for just 3 %.
Figure 2 compares the blue area (theoretical capability) to the red area (observed exposure) across occupational categories. For example, Claude covers only 33 % of tasks in the Computer & Math category despite 94 % theoretical feasibility.
The gap indicates substantial room for future AI adoption.

Most and least exposed occupations

Key point: Computer programmers, customer‑service reps, and data‑entry keyers rank highest, while many service and manual jobs have near‑zero exposure.

Top‑10 occupations (Figure 3) show coverage ranging from 75 % (Computer Programmers) to 67 % (Data Entry Keyers).
Approximately 30 % of workers have zero observed exposure; examples include Cooks, Motorcycle Mechanics, Lifeguards, Bartenders, Dishwashers, and Dressing‑Room Attendants.

Exposure and projected job growth

Key point: Occupations with higher observed exposure have slightly weaker BLS employment growth projections for 2024‑2034.

A weighted regression finds that a 10‑point increase in observed exposure corresponds to a 0.6‑point reduction in projected growth.
No significant correlation appears when using the theoretical β metric alone.

Worker demographics by exposure level

Key point: Workers in the top exposure quartile tend to be older, female, more educated, higher‑paid, and more likely to be Asian or white.

Compared to the zero‑exposure group, the high‑exposure group earns 47 % more on average.
Graduate‑degree holders constitute 17.4 % of the high‑exposure group versus 4.5 % of the low‑exposure group.

Unemployment trends for exposed workers

Key point: Since the release of ChatGPT, unemployment rates for highly exposed workers have not risen relative to low‑exposure workers.

Figure 6 shows parallel unemployment trends for the top‑quartile exposure group and the zero‑exposure group after the COVID‑19 shock.
Difference‑in‑differences analysis yields a small, statistically insignificant increase in the unemployment gap.
The framework can detect differential unemployment changes of roughly 1 percentage point; observed effects are well below this threshold.

Hiring dynamics for young workers

Key point: Job‑finding rates for workers aged 22‑25 have fallen modestly in high‑exposure occupations, suggesting a slowdown in hiring rather than increased layoffs.

Figure 7 shows the monthly job‑finding rate dropping from ~2 % to ~1.5 % in high‑exposure jobs after 2024, while low‑exposure jobs remain stable.
The post‑ChatGPT average estimate is a 14 % decline in hiring for exposed occupations, barely reaching statistical significance.
No comparable decline is observed for workers older than 25.

Interpretation and limitations

Key point: Early evidence points to limited direct displacement but hints at emerging hiring frictions for younger workers in AI‑exposed roles.

The analysis focuses on unemployment because it directly reflects economic harm; other metrics (job postings, occupational mix) may capture different dynamics.
Results depend on the chosen exposure threshold; varying the cutoff from median to 95th percentile does not produce a positive unemployment effect.
Survey‑based measures (CPS) may under‑capture job transitions, especially for recent graduates or those exiting the labor force.

Future work

Key point: Anthropic plans to update the exposure metric with newer usage data, refine the theoretical β scores, and investigate outcomes for recent graduates in exposed fields.

Incorporating additional platforms and international data will broaden the applicability of the framework.
A deeper look at hiring pipelines for AI‑related degrees could clarify the early signals observed for young workers.

The full technical appendix is available here.