When AI Exposure Scores Disagree by 11x, a New Statistical Fix Steps In

A factor-of-eleven gap between competing measures of "AI exposure" sounds like a rounding error until you realize it can flip the sign of an economic conclusion. That is precisely the problem tackled in a new paper submitted to arXiv on July 13, 2026, which proposes a statistical method to reconcile wildly inconsistent occupational AI-exposure scores and still recover a reliable structural estimate of how automation relates to employment.

The core issue is one familiar to applied economists: researchers have built several different scores to estimate how exposed a given occupation is to artificial intelligence, drawing on tools like large language models and Webb's patent-text methodology. These scores are meant to measure the same underlying concept, yet they diverge so sharply that using one instead of another can change the sign of an estimated employment effect. The paper documents exactly this: the coefficient linking AI exposure to post-2022 employment flips sign depending on whether it is built from language-model-based measures or the Webb patent-text measure.

Rather than picking a winner among the competing scores, the authors treat the problem as one of latent-variable econometrics. They model the true, unobserved regressor (actual AI exposure) as something observed only through multiple noisy, nonlinear measurements, essentially a classic errors-in-variables setup with several imperfect proxies standing in for one real signal. Their fix requires the "consensus" measurement function tying these proxies together to be linear, and it imposes bounds on how much the curvature (nonlinearity) of each individual source's error can differ from the others. That structural discipline is what allows them to pin down the relationship despite the noisy, disagreeing inputs.

The payoff is a closed-form interval for the true structural coefficient, centered on a symmetric estimator built across all the sources. When the method is applied to five retained exposure measures, it produces a loading-invariant consensus coefficient of -0.239. Critically, the resulting partial-identification interval is tight: its half-width is just 1.23 percent of the point estimate, meaning the sign-flipping ambiguity seen in raw measures effectively disappears once the correction is applied.

Validating the Approach at Scale

The authors test their framework against a substantial real-world dataset: an American Community Survey panel covering 8.88 million person-year observations spanning 2015 to 2024. This scale lets them check whether their identification strategy holds up outside of simulation, applying it directly to the employment-exposure puzzle that motivated the paper. For cases with at least four available measurements, they show the bound can be estimated using a split-instrument auxiliary regression paired with Imbens-Manski confidence intervals, a technique standard in partial-identification econometrics for handling intervals rather than single point estimates.

For labor economists and AI policy researchers, the practical implication is straightforward: the choice of which AI-exposure score to use has been a hidden source of contradictory findings in the literature. This paper offers a way to extract a stable, defensible estimate even when the underlying measurement tools themselves cannot agree.