AI Exposure Scores: what they measure, what they miss, and what comes next

A new research paper critiques the widely-cited GPT exposure scores from 2023, which measure how many occupational tasks AI can assist with, revealing critical gaps between static measurements and dynamic policy needs. The authors identify a structural measurement problem and a deeper coordination failure between researchers and policymakers, proposing frameworks that incorporate temporal dynamics, worker perspectives, and actual adoption data to better inform AI workforce policy.

The GPT exposure scores produced by Eloundou et al. in 2023 have become foundational to policy discussions about AI's labor market impact, yet their limitations haven't kept pace as these metrics circulate beyond their original context. The paper reveals how static measures fail to capture the temporal, geographic, and conceptual nuances required for effective policymaking. As exposure scores travel from academic papers into policy briefs and public discourse, their caveats get lost, creating a widening gap between what researchers actually measured and what policymakers need to know. The authors identify five emerging research families addressing these gaps: dynamic measures that track change over time, ensemble methods combining multiple data sources, task-framework extensions, worker-centered metrics prioritizing human perspectives, and adoption data reflecting real-world deployment patterns. However, the more critical problem the paper highlights is the coordination breakdown itself. Policymakers continue citing static scores without engaging with methodological innovations that would substantially improve reliability. The paper calls for a bidirectional commitment: researchers must build better data infrastructure, adopt participatory methods including workers as knowledge partners, and write explicitly for policy audiences; simultaneously, policymakers must diversify their evidence base, shift from predicting futures to preparing for multiple scenarios, and treat workers as epistemic partners rather than passive subjects. The analysis underscores that improving AI labor impact assessment requires not just better metrics but fundamentally restructured communication and collaboration between technical and policy communities.

• →GPT exposure scores, widely used in policy debates, lack temporal and geographic specificity needed for actionable AI workforce planning.
• →Five research families now address measurement gaps through dynamic benchmarks, ensemble methods, task extensions, worker-centered approaches, and adoption data.
• →The primary problem is not measurement alone but coordination failure between researchers publishing updates and policymakers still citing original static scores.
• →Effective AI policy requires shifting from prediction-focused models to preparedness frameworks that engage workers as knowledge partners.
• →Closing the research-policy gap demands bidirectional change: researchers writing for policy audiences and policymakers widening their evidence base.