Designing Human-AI Interactions for Improvement and Objective Alignment.

When AI systems are deployed, human users often develop beliefs about their operation, which in turn influences how they interact with the system, potentially leading to strategic behaviors or "gaming." This creates a dynamic human-AI interaction loop where the system's feedback or decisions shape user actions. This research focuses on designing these interactions to achieve three critical goals. Firstly, the aim is to help individuals accurately understand how AI systems function, enabling them to improve their performance or secure favorable outcomes efficiently. Secondly, the design seeks to foster genuine improvement behaviors while discouraging manipulative or gaming tactics. Finally, the research ensures that the AI system continues to meet its core objectives, such as maximizing accuracy, despite these human interactions. The thesis employs a multi-faceted methodological approach, combining theoretical analysis, data-driven modeling, human-subject experiments, and empirical evaluations using both real-world and synthetic datasets. This comprehensive work contributes to human-centered machine learning by providing principles and methods for creating AI systems that are aligned with human needs, values, and capabilities, ultimately leading to more effective and ethical AI deployments.