Abstract

This paper investigates the integration of Artificial Intelligence (AI) in higher education, exemplified by the KI4Edu project. It responds to students' growing expectations for digital, flexible, and personalized learning, alongside increasing institutional challenges such as staff shortages and rising demand. Large Language Models (LLMs) offer transformative potential but also introduce risks, including hallucinations, opacity, and limited contextual accuracy. This study presents a pedagogical framework and technical architecture for responsible AI deployment. Educators assume the role of mentors, ensuring quality assurance, critical reflection, and content curation, while AI supports knowledge accessibility, structuring, and scalability. The approach is implemented through a Retrieval-Augmented-Generation (RAG) system that consolidates verified teaching materials within a knowledge database, with an avatar serving as an interactive interface for personalized learning pathways. Key contributions include: (1) a systematic analysis of challenges in AI-assisted education, (2) a pedagogical model redefining educators' roles, and (3) a technical concept for avatar-based knowledge transfer. The paper discusses limitations, such as LLM output variability and social risks like "AI Psychosis," and highlights the need for monitoring mechanisms, scalable system architectures, and a balanced integration of technological innovation with human responsibility in teaching.