PILAR: Personalizing Augmented Reality Interactions with LLM-based Human-Centric and Trustworthy Explanations for Daily Use Cases

AI explanations in augmented reality need to be personalized and real-time: Traditional explainability methods struggle to provide dynamic, context-specific insights that adapt to individual users' needs in everyday AR applications, highlighting a gap between current AI transparency approaches and what users actually experience in real-world systems.

Large language models can unify fragmented explanation approaches: Rather than using separate techniques to explain different aspects of AI decisions (when, what, and how), PILAR demonstrates that a single LLM-based framework can generate cohesive, human-centered explanations tailored to each user's preferences and circumstances.

Personalized explanations build trust in AI-powered AR systems: By adapting explanations to individual users' needs—such as dietary preferences in the recipe recommendation prototype—the framework aims to foster greater user trust and engagement with AI-driven augmented reality tools in daily life.

Real-world implementation matters for responsible AI: The paper moves beyond theoretical explainability by implementing PILAR in a practical AR application with real-time object detection and personalized recommendations, demonstrating that responsible AI requires testing explanations in actual use cases rather than laboratory settings alone.