Website CV Paper
Fields
Applied Micro, Innovation, Public, Industrial Organization, Econometrics
Job market paper
When More Precise Algorithmic Predictions Meet Biased Human Discretion
I develop a model of unbiased predictions mediated by biased humans to show that a bias-neutral precision gain in predictions (e.g. from AI adoption) is not generally bias-neutral in its effects. Expected victims of bias are discriminated-group individuals with latent risks meeting an unbiased standard but failing the stricter, biased standard they face. For them, precision raises discrimination by reducing noise-driven chance approvals. For other disadvantaged-group individuals, finite precision gains can either increase or decrease discrimination, although discrimination against them vanishes in the noise-free limit. Consequently, average discrimination can increase or decrease. Exact neutrality requires a knife-edge cancellation. A sizable increase in average discrimination is possible when baseline predictions are sufficiently noisy. Opposing effects can cancel to first order otherwise, leaving a second-order change. I structurally estimate the mechanism and test its predictions using Virginia felony sentencing data. In a structural counterfactual holding prediction-side bias and human bias fixed, a precision gain raising predictive accuracy marginally (AUC from 0.586 to 0.596) raises discrimination against expected victims by 7 percentage points while average discrimination increases only slightly. Thus, average discrimination can mask sizable distributional effects as more precise predictions concentrate discrimination on expected victims. This shows a new, precision-driven channel of AI potentially exacerbating discrimination, distinct from biased model design and biased training data.