Since the rapid proliferation of Large Language Models (LLMs), the majority of university students report using Generative Artificial Intelligence (GenAI) in their studies. While GenAI offers benefits for personalized learning, it raises ethical, legal, and societal concerns, including issues of accuracy, bias, and uncritical use. AI literacy is increasingly recognized as a core competency and is now embedded in policy frameworks such as the EU AI Act. However, current educational approaches often emphasize conceptual knowledge or isolated technical skills, with limited evidence on how to develop AI literacy as an applied, practice-based competency. This study investigates a practice-based AI literacy intervention within a Master’s programme in Medical Informatics. Grounded in constructionist learning theory, the intervention involves a week-long project in which students design, build, and validate GenAI-based tutors using a secure LLM platform. The “users-to-makers” approach enables students to translate ethical, technical, and pedagogical principles into concrete design decisions. A mixed-methods pre-post study was conducted with 43 students, using rubric-based assessments, self-reported competency surveys, and a 3-month follow-up to assess behavioral transfer. Results show strong gains in technical competencies and moderate improvements in applying ethical principles. Students reported increased self-efficacy, although bias mitigation remained a persistent challenge. A key finding was a discrepancy between perceived and demonstrated competence, with students expressing confidence in responsible AI use while struggling to implement it effectively. Practice-based, constructionist approaches enable the integration of technical, ethical, and pedagogical competencies while revealing gaps between understanding and application. This study provides an evidence-based, transferable model for AI literacy education, including a validated assignment structure and assessment framework.