AI Exposure Scores Cannot Yet Guide Labor Policy, Researchers Warn

Policymakers relying on AI exposure estimates face significant blind spots

As artificial intelligence spreads across industries, researchers have developed AI exposure scores to estimate which jobs and tasks the technology might affect. These calculations have become popular tools for understanding potential labor market disruption, but they come with serious limitations that make them unsuitable for guiding policy decisions on their own.

AI exposure scores attempt to quantify how much a given occupation might interact with artificial intelligence. Researchers calculate these scores using methods ranging from expert assessments to having AI models evaluate their own capabilities, plus analysis of job descriptions and chatbot usage rates. While these approaches fill a critical data gap, they remain approximations built on assumptions rather than direct measurements of actual AI deployment and impact.

Why it matters

Policymakers designing worker support programs or economic interventions need accurate information about how AI is reshaping employment. Using exposure scores as if they were hard data could lead to misdirected resources or policies that address phantom problems while missing real disruptions. Understanding what these estimates can and cannot reveal is essential for effective labor market policy in an era of rapid technological change.

The isolation problem

Research using AI exposure scores has produced conflicting results, with studies finding positive, negative, and neutral employment effects. This variation partly reflects different calculation methods, but a more fundamental problem affects all such research: the inability to separate AI's impact from other economic forces.

One prominent study found that AI exposure correlated with stagnant or declining employment for early-career workers. Yet the authors themselves noted they could not definitively attribute this pattern to AI, acknowledging in their February 2026 update that employment declines before 2024 likely stemmed from broader macroeconomic factors.

The California Unemployment AI Tracker, developed with the California Employment Development Department and California Policy Lab, observed increased unemployment insurance claims among highly educated workers and those in San Francisco's tech sector who had high AI exposure. The tool's creators emphasized that these patterns could reflect various economic influences beyond AI adoption.

Similarly, Yale Budget Lab analysis found no clear evidence that large language models like ChatGPT were affecting U.S. employment or wages, while noting that the findings' reliability depends on successfully distinguishing AI introduction from other labor market shocks.

Methodological concerns

Beyond empirical limitations, the quality of AI exposure scores depends entirely on the assumptions and methods used to create them. A working paper from researchers at Northwestern University and American University revealed that scores calculated using large language models fluctuate substantially depending on which AI model performs the evaluation. Any bias in the AI system used to generate exposure scores carries through to distort subsequent analysis.

This finding highlights a deeper issue: these scores represent predictions about AI's interaction with work, not verified measurements. Without actual data on AI deployment and outcomes to validate against, researchers cannot confirm whether current exposure estimates accurately reflect reality.

The path forward

For policymakers to craft effective responses to AI-driven labor market changes, they need measurements of actual AI use and impact rather than estimates. AI exposure scores can suggest where to look for potential disruption and help target initial research, but they cannot substitute for comprehensive data on how firms implement AI and how workers experience these changes.

These details were first reported by the Washington Center for Equitable Growth in an analysis published in August 2026. The organization maintains an ongoing database tracking AI and labor market research.