In the midst of the controversy surrounding its severed deal with the US Department of War, Anthropic has released a new study on the effects of AI on jobs and the labour market. The study, called “Labor Market Impacts of AI: A New Measure and Early Evidence,” reveals how the current job market is affected by the use of AI. However, the report doesn’t give a fair idea of everything the data suggests and here’s why.
Anthropic’s economists introduced “observed exposure” — a metric combining theoretical capabilities of large language models (LLMs) with real-world usage data from millions of Claude conversations and API calls. The study weighted automated uses more heavily than assistive ones, focused on work-related activity, and mapped interactions to standardised O*NET job tasks.
What the Anthropic study found
Key findings included no systematic rise in unemployment for highly exposed workers since late 2022 (when tools like ChatGPT launched), though hiring of younger workers (ages 22-25) appeared to slow in those roles. Occupations with higher exposure, often more educated, higher-paid, female-dominated, and older, were projected by the US Bureau of Labor Statistics (BLS) to grow more slowly through 2034.
A Forbes report now contends that the study’s foundation is too narrow to support broad claims about economy-wide effects. The core data comes solely from Claude interactions, reflecting Anthropic’s specific user base rather than widespread AI adoption across tools like ChatGPT Enterprise, Microsoft Copilot, or Google Gemini.
Key criticisms of the Anthropic study
In the report, there are three major identified problems:
Data only based on Claude usage:
The study tracks only Claude usage, potentially biased toward Anthropic’s demographics and access patterns. Most workplaces rely on other platforms with varying levels of integration with the workflow. Distinguishing work from personal use relies on platform signals that can be imperfect, and API traffic, which is weighted as deeper integration, might include testing rather than true production deployment. Minimum usage thresholds could classify early, low-frequency adoption as “zero exposure.”
Mapping and Interpretation
Prompts like “write a client email” might apply to multiple occupations (sales, HR, legal), leading to fuzzy classifications. Observed usage reflects barriers like permissions or reliability concerns, not just task suitability. Capability baselines use outdated estimates, ignoring rapid model improvements and real-world constraints like verification needs.
Causal attribution
The study focuses on unemployment as a blunt indicator, potentially missing subtler shifts such as reduced hiring, slower promotions, fewer entry-level roles, or wage compression. Comparisons between exposed and less-exposed jobs remain vulnerable to external factors like economic cycles or trade policies. Even productivity gains might expand output (via the Jevons Paradox) rather than cut jobs.
The report from Forbes stresses that assistive AI can still displace labour demand indirectly, while “automated” uses often require human oversight. The reported gap between capability and usage, framed as underdeployment, is sensitive to methodological choices and does not prove widespread risk.