Who Uses AI? Platforms, Workforce, and AI Exposure

2026-05-23 · Source: cs.AI updates on arXiv.org · Field: Science & Research — Mathematics & Computational Sciences, Social Sciences & Behavioral Studies, Economic Analysis & Policy · Depth: Expert, extended

Summary

A study by Yin and Ogut reveals that AI occupational exposure measures derived from platform conversation logs are significantly biased, primarily reflecting the platform's user base rather than the actual workforce composition. Their analysis shows that varying the platform input (e.g., Anthropic Claude consumer vs. enterprise, Microsoft Copilot) can change the post-ChatGPT employment coefficient by a factor of 1.9, with estimates sometimes disagreeing in sign. Reweighting these platform shares to Bureau of Labor Statistics (BLS) workforce demographics attenuates employment estimates by 42% to 93%. The authors formalize this as non-classical measurement error, deriving partial-identification bounds and demonstrating that this bias systematically understates AI's substitution effects more than its augmentation effects, leading to inconsistent findings in the AI-and-labor literature and potentially misdirecting policy resources.

Key takeaway

For research scientists and data scientists analyzing AI's labor market impacts, you must critically evaluate platform-derived exposure measures. Your studies should reweight these measures to Bureau of Labor Statistics workforce shares and report partial-identification bounds, not just point estimates, to account for user-base bias. Policy makers allocating resources for AI retraining should avoid relying solely on platform-weighted exposure data, as it systematically misdirects funds away from vulnerable, underrepresented populations.

Key insights

AI exposure measures from platform data are biased by user demographics, not true workforce exposure, distorting labor market impact estimates.

Principles

• AI platform user bases are not representative of the general workforce.
• Measurement error in AI exposure is non-classical and platform-specific.
• Bias understates AI's substitution effects more than augmentation.

Method

Formalize platform-selection bias as non-classical measurement error. Reweight platform conversation shares to BLS workforce shares to remove between-occupation bias and derive partial-identification bounds for employment elasticity.

In practice

• Reweight platform-derived exposure measures with workforce demographics.
• Report partial-identification intervals, not single point estimates.
• Recognize platform data overrepresents high-skill, high-income users.

Topics

• Artificial Intelligence
• Labor Markets
• Measurement Error
• Occupational Exposure
• Non-Probability Sampling

Best for: AI Scientist, Research Scientist, Data Scientist, Policy Maker

Related on AIssential

• US starting to witness heavy job losses in occupations exposed to AI - Business Standard — AI deployment is beginning to reshape US employment patterns, particularly in exposed occupations.
• #357 Data-Driven Workforce Analytics with Ben Zweig, CEO at Revelio Labs — Effective labor markets require standardized, granular job taxonomies, defining jobs as task bundles, not skills.
• The Interview Is Broken. And Most AI Is Making It Worse. — Unstructured interviews are inherently unreliable, and AI often amplifies existing biases rather than improving hiring validity.
• SpaceX IPO Multiple Times Oversubscribed — The AI and space sectors are seeing massive capital inflows, but also significant risks and labor market shifts.
• Mar 5, 2026Economic ResearchLabor market impacts of AI: A new measure and early evidence — AI's labor market impact is better understood by combining theoretical capability with observed, automated usage.
• Predicting AI job exposure — Predicting AI's impact on jobs is inherently complex and cannot be accurately quantified by static job descriptions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.