The rapid advancement of artificial intelligence (AI) has produced remarkable achievements, from generating realistic text and images to designing new proteins and supporting healthcare delivery. However, the opaque nature of these powerful technologies poses significant challenges to scientific progress and public trust. When AI systems make predictions in sensitive domains such as medicine, education, or public policy, we need to understand which factors drive their decisions and whether their discoveries are reliable and reproducible. This project develops methods to extract interpretable, verifiable and trustworthy insights from sophisticated AI algorithms. By doing so, it will accelerate scientific discovery across disciplines, while strengthening public confidence in data-driven research. Beyond methodological innovation, the project will contribute to training the next generation of AI researchers and data scientists. This project will develop novel statistical methods for testing large number of hypotheses about variable importance in complex predictive models, with rigorous control of false discovery rates. The research will focus on context-dependent variable importance, studying how explanatory variables influence outcomes under different conditions and through non-additive interactions, leveraging sophisticated machine learning architectures. The methodological framework will integrate recent advances including knockoff inference, e-values, conditional randomization tests, and explainable AI techniques. A key innovation will be designing inference procedures robust to multiple data passes, avoiding selection bias and circular reasoning, while adapting to the signal in the data. Applications will focus on genomic data analysis, including inference on gene-gene interactions. The project will contribute to training one graduate student. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Fou