This action research study investigates how video lectures enhance engagement and transform learning attitudes among 42 Grade 4 EFL students at Thuy Duong Primary School. Conducted over 12 weeks, the research compared traditional linear YouTube cartoons in Cycle 1 with interactive, AI-generated video lectures featuring local school contexts and virtual teacher avatars in Cycle 2. Data collection was triangulated using classroom observation checklists, student questionnaires, and a teacher’s reflective journal. Quantitative analysis revealed a statistically significant improvement in student engagement across behavioral, emotional, and cognitive dimensions, alongside a positive shift in learning attitudes. Qualitative findings confirmed that interactive AI features and pause-and-predict techniques effectively lowered the learners' affective filters, transitioning them from passive consumers to active speakers. The study implies that teachers should transition from passive consumers of generic online content to active designers of customized, locally relevant digital resources. Utilizing AI to personalize learning environments drastically minimizes extraneous cognitive load and cultural barriers for young learners. Furthermore, embedding structured interactive pauses transforms standard viewing into generative processing , while virtual avatars provide a psychologically safe scaffolding space that builds communicative confidence before students engage in real-world interactions