Abstract Generative AI tools like Google Gemini have created new opportunities for language learning. However, empirical evidence regarding their impact on oral proficiency remains limited. This quasi-experimental study investigates the effectiveness of Google Gemini as a scaffolding tool for Vietnamese university freshmen (N=78). Over a seven-week intervention, the experimental group utilized Gemini for ideation, real-time feedback, and conversation simulation, while the control group followed traditional instruction. Data was collected via pre- and post-tests based on the CEFR B1 rubric, alongside anxiety surveys and semi-structured interviews. Paired sample t-tests revealed that the AI-assisted group demonstrated statistically significant improvements in Fluency and Discourse Management (p < .05) and a substantial reduction in speaking anxiety compared to the control group. However, no significant differences were observed in Grammatical Accuracy or Lexical Resource, suggesting that while AI effectively lowers the affective filter and promotes communicative flow, it may require longer-term integration to enhance linguistic precision. These findings challenge the view of AI as a mere correction tool, proposing instead its role as a psychological scaffold that empowers reticent learners to speak more confidently.