Who Uses AI? Platform Selection and the Measurement of Occupational AI Exposure

Researchers demonstrate that AI exposure measurements derived from platform conversation logs significantly misrepresent actual occupational AI adoption across the workforce. The study reveals that platform-based metrics conflate AI task applicability with user demographic composition, producing estimates that vary by 90% depending on data source and can even reverse directional findings about AI's employment impact.

This research addresses a critical methodological flaw in how economists and policymakers currently measure AI's workplace penetration. As AI adoption accelerates, researchers increasingly rely on conversation logs from platforms like ChatGPT to estimate which occupations face disruption or augmentation. However, these logs capture early adopters and tech-forward professionals rather than representative workforce samples, creating substantial measurement bias.

The findings are striking: using different platform data sources changes employment coefficient estimates by factors exceeding 1.9x, and consumer versus enterprise channels within the same vendor sometimes produce opposite conclusions about substitution effects. This divergence stems from selection bias operating at two levels—between occupations and within them—where different user types gravitate toward different platforms based on adoption barriers, pricing, and industry culture.

The implications extend beyond academic precision. Policymakers considering AI regulation, education investment, or workforce retraining programs rely on occupational exposure metrics to prioritize interventions. If these metrics overstate disruption risk in high-visibility occupations while understating it elsewhere, policy responses become misdirected. The authors' reweighting adjustment—using Bureau of Labor Statistics employment shares—reveals that conventional estimates overstate AI exposure by 42-93 percent.

The research also reframes the AI impact narrative. Current platform data better captures how observed users augment their capabilities rather than how widespread substitution actually occurs in the broader economy. This distinction matters for investor confidence and labor market stability assessments. Going forward, researchers must triangulate multiple data sources and explicitly account for selection mechanisms rather than treating platform logs as workforce proxies.

• →AI exposure measurements from platform logs overstate actual workforce exposure by 42-93% when reweighted to representative employment distributions.
• →The same AI vendor's consumer and enterprise platforms produce contradictory conclusions about AI's employment effects, revealing severe platform-selection bias.
• →Current metrics capture augmentation among early adopters better than substitution risk across the broader workforce.
• →Policymakers relying on platform-derived exposure scores risk directing workforce retraining and regulation efforts toward incorrect occupational priorities.
• →Measurement error from platform selection is non-classical and requires specialized econometric approaches to construct valid confidence bounds.