Teach AI What It Doesn't Know

Abstract AI systems are rapidly transitioning from laboratory demonstrations to decision‐making technologies deployed in high‐stakes domains. Yet reliability remains a primary obstacle to responsible adoption: discriminative models can be confidently wrong under out‐of‐distribution (OOD) inputs, and foundation models (FMs) such as large language models (LLMs) can generate fluent but untruthful, harmful, or misaligned outputs. My research develops the foundations of reliable machine learning with minimal human supervision , unifying algorithms, and theory that make reliability a first‐class objective alongside accuracy. I advance unknown‐aware learning through automated outlier generation, introducing feature‐ and input‐space synthesis frameworks that regularize decision boundaries and improve interpretability. I further establish principled methods for learning “in the wild” by leveraging unlabeled deployment data under mixture and contamination models, with theoretical guarantees and state‐of‐the‐art performance for OOD detection and generalization under diverse shifts. Finally, I design reliability frameworks for FMs by exploiting unlabeled signals to detect hallucinations, defend against malicious prompts in vision–language models, and denoise noisy preference data for more dependable alignment. Collectively, these contributions provide a cohesive toolkit for deploying AI systems that remain accurate, calibrated, and trustworthy in open‐world environments.