Equipping in-service teachers to teach Artificial Intelligence (AI) and Machine Learning (ML) meaningfully demands more than technical instruction: it requires course designs that connect unfamiliar concepts to existing pedagogical knowledge and classroom practice. This experience report describes the design and retrospective evaluation of professional development courses delivered to 42 in-service teachers across two contexts, grounded in the Technological Pedagogical Content Knowledge (T-PACK) framework and Bloom’s Taxonomy. We report on hands-on learning activities in which participants generated, labeled, and evaluated their own training data — for instance by training a classifier on photographs of their own facial expressions or by teaching a browser-based game to recognize their body gestures — and reflect on why this data-centric, embodied approach seems to be effective at building conceptual understandings. A recurring theme is the value of connecting new AI content to subject-area knowledge teachers already hold, particularly from media education, rather than treating AI as a wholly separate domain. We also identify a gap in available curriculum materials: resources that explain algorithmic bias not merely as an observable problem but in terms of its technical origins in data collection, labeling, and model training remain scarce. Findings are discussed with reference to embodied cognition, body syntonicity, and experiential learning, and recommendations are offered for practitioners designing AI professional development for teaching staff.