Machine learning models are increasingly deployed in real-world settings where distribution shifts—differences between training and deployment data—can significantly impact their reliability. These shifts affect models in multiple ways, leading to degraded generalization, fairness collapse, loss of robustness, and new safety vulnerabilities. This dissertation investigates how to build reliable AI under distribution shifts, providing theoretical insights and practical solutions across diverse applications.

We begin by studying generalization under distribution shifts, exploring how model invariance affects performance. We introduce a theoretical framework that quantifies the role of data transformations in shaping generalization, providing insights into selecting transformations that improve model robustness in shifted environments. This foundation also extends to fairness, where we examine how pre-trained fair models fail when deployed in new distributions and propose a method to transfer fairness reliably under distribution shifts.

Next, we focus on robust perception and AI-generated content under shifting distributions. We investigate how models interpret visual information, showing that contextual reasoning can help mitigate spurious correlations and improve robustness under domain shifts. We also assess the reliability of AI-generated content, revealing how image watermarks, designed for provenance tracking, often fail when subjected to real-world distortions and adversarial attacks. To address this, we introduce a comprehensive benchmark for evaluating watermark robustness, providing a framework for improving their reliability.

Finally, we turn to safety in large language models (LLMs) and investigate how distribution shifts in training and deployment introduce new vulnerabilities. We analyze false refusals in safety-aligned LLMs, demonstrating that misaligned decision boundaries lead to excessive conservatism at test time. We also explore retrieval-augmented generation (RAG) models, showing that despite their promise, they can introduce new safety risks when deployed in settings for which they were not originally trained. Our findings highlight critical gaps in existing AI safety evaluations and emphasize the need for new methods tailored to evolving AI architectures.

By addressing generalization, robustness, and safety under distribution shifts, this dissertation contributes to a deeper understanding of these challenges and provides practical strategies for improving AI reliability in real-world deployment.