Modern artificial intelligence systems produce striking images, text, and other creative outputs, but it is often unclear what these systems have actually learned internally. This makes it difficult to ensure that these models are reliable, safe, and trustworthy when deployed in the real world. Although these models can imitate patterns in data, the process through which they do so does not necessarily correspond to meaningful causes, stable mechanisms, or interpretable concepts that stakeholders can decipher and diagnose. This project develops a new statistical framework for building AI models designed to uncover interpretable, generalizable structures hidden inside complex, high-dimensional data such as images, language, and scientific measurements. This research investigates the statistical foundations of generative AI. A key goal is to understand how and when generative models learn reusable, causal structure and what the tradeoffs are. Specifically, the project focuses on understanding how generative models learn complex, high-dimensional structures without suffering the curse of dimensionality and how they can learn interpretable causal factors from data. This will deliver a framework with practical models and algorithms for reliable generative AI that is both independently verifiable and reproducible. The work will combine ideas from causal inference, nonparametric statistics, latent variable modeling, and deep learning to develop methods with rigorous guarantees. The goal is to move beyond black-box imitation toward AI systems whose internal factors can be interpreted, tested, and used to understand how complex systems change. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.