Towards Trustworthy Statistical Inference with Black-Box AI Predictions

The widespread adoption of AI and ML has reshaped modern data analysis. Predictions, embeddings, and synthetic data from black-box models, such as deep neural networks and large language models, are increasingly incorporated into downstream statistical workflows. For instance, predicted gene expression values or polygenic risk scores are often substituted for experimental assays, enabling researchers to enlarge cohorts and pursue hypotheses when direct measurement is infeasible, costly, or time-consuming. While AI/ML models can usually deliver strong predictive performance, their opaque mechanisms and potential biases introduce additional layers of uncertainty that can compromise the validity of classical inference. Treating black-box outputs as ground truth risks biased estimation, misleading confidence intervals, and invalid hypothesis tests. This course offers a principled overview of statistical challenges and emerging solutions for trustworthy inference. We cover methods that explicitly account for prediction-related uncertainty, improving both validity and efficiency when analyzing data that blend observations with AI-derived outputs. Applications from biomedical and social sciences will illustrate not only the risks of naïve reliance on AI predictions but also the opportunities created by rigorous statistical integration. The course is designed for statisticians, data scientists, and applied researchers seeking to leverage AI predictions in their work without compromising inferential integrity.