In the digital age, academic success increasingly depends on information problem solving (IPS) competence: the ability to find, evaluate, process, and synthesize information into knowledge products. However, generative AI (genAI) transforms how higher education students engage with digital information, challenging models developed for web-based IPS, such as the IPS using Internet (IPS-I) model. Therefore, an empirically grounded model for the AI-mediated IPS process is needed. This study decomposes the cognitive activities of the AI-mediated IPS process and pinpoints where students need guidance and support. It empirically refines the IPS-I model through cognitive task analysis of think-aloud protocols from 14 novices (pre-service teachers) and 12 experts (teacher trainers) completing an authentic IPS task using genAI. Through bottom-up analysis of observed behaviors, we identified 25 cognitive activities distributed across six interrelated phases: *Define*, *Search*, *Select*, *Process*, *Synthesize*, and *Create*. Compared to the original IPS-I model, three new and four modified activities emerged, with Regulation remaining a central, cross-cutting metacognitive component. Additionally, experts and novices differed in time allocation across phases, AI tool usage, and sequencing of the IPS phases. Experts exhibited a more iterative process, integrating AI across all phases, including Regulation, and delayed the use of AI. In contrast, novices used AI early, mainly during *Search* and *Create*. The resulting Digital Information Problem Solving (DIPS) model provides a refined, empirically grounded framework for understanding and teaching IPS within the context of genAI. This model defines the intended learning outcomes for learning environments aimed at fostering responsible AI use in higher education.
