Chat as Learning in Interactive Information Retrieval and Generation

Generative AI-based conversational systems are rapidly becoming the default gateway to information, yet their educational potential and risks remain under-examined. This paper introduces Chat as Learning (CaL), a paradigm that views dialogue not as a veneer over search but as the central mechanism for knowledge construction. We present a three-layer framework that couples fine-grained learner modeling with adaptive prompting support and a retrieval-augmented generator that blends the model's prior knowledge with verified, up-to-date sources. Each conversational turn functions as a diagnostic probe, which enables real-time refinement of intent, mastery level, and motivational state. Building on this architecture, we map a cross-disciplinary research agenda that spans user models, reinforcement-optimized prompt policies, information-quality and knowledge gap signaling, multimodal assessment instruments, and bias mitigation. By foregrounding both the promise of democratized expertise and the threat of over-reliance on fluent yet fallible outputs, CaL establishes an urgent agenda for reproducible benchmarks and responsible deployment of conversational learning systems.