While AI has become more prevalent in our society than ever, many young learners are found holding various naive, erroneous conceptions of AI due to the influence of their technology and media environments. To address this issue, this study seeks to propose a novel pedagogical solution to improve upper-elementary school students' scientific understanding of AI. Following a theory-informed design convention, we propose an embodied, analogical and disruptive (EAD) approach which is centred on a human-AI comparison through analogical teaching and embodied interaction. To evaluate the impact of this approach, a matched-group experimental study with pre- and posttest interviews was conducted among 77 Grade 6 elementary students in China. The statistical analysis showed that the experimental group, learning via the EAD approach, significantly outperformed their counterparts receiving direct instruction in terms of student growth in understanding. Qualitative analyses revealed the strengths of the EAD approach in supporting student engagement, abstract thinking and system thinking, and its limitations in cognitive overload and communication issues. Patterns of student learning were identified, including their mental schemas and strategies. The EAD approach, as an evidence-based, age-appropriate pedagogical solution, demonstrates the value of embodied cognition and human-AI analogy in AI education in elementary education.