Abstract
We introduce prediction-powered inference $\unicode{x2013}$ a framework forperforming valid statistical inference when an experimental data set issupplemented with predictions from a machine-learning system such as AlphaFold.Our framework yields provably valid conclusions without making any assumptionson the machine-learning algorithm that supplies the predictions. Higheraccuracy of the predictions translates to smaller confidence intervals,permitting more powerful inference. Prediction-powered inference yields simplealgorithms for computing valid confidence intervals for statistical objectssuch as means, quantiles, and linear and logistic regression coefficients. Wedemonstrate the benefits of prediction-powered inference with data sets fromproteomics, genomics, electronic voting, remote sensing, census analysis, andecology.