Problem-Based Learning with GenAI: Effects on Students' Statistical Thinking

Abstract

The purpose of this study was to determine the effects of a problem-based learning (PBL) environment, with and without the support of generative artificial intelligence (GenAI), on senior high school students’ statistical thinking. An embedded mixed-methods quasi-experimental design was employed involving three intact Grade 11 STEAM classes randomly assigned to GenAI-PBL, PBL, and conventional groups. Over 12 weeks, students studied four statistics lessons through either a PBL cycle or lecture-discussion format. Data sources included the Statistical Thinking Test (TST), end-of-lesson evaluations, daily journals, the Course Experience Questionnaire (CEQ), focus group discussions, and ChatGPT conversation logs. Results showed no significant differences in overall TST scores among groups, although patterns emerged in specific subtasks. The GenAI-PBL group performed relatively better in designing sampling plans, the PBL group excelled in justifying statistical tests, and the conventional group demonstrated steady procedural improvement over time. Subtask E, explaining the purpose of hypothesis testing, consistently received the lowest scores across groups. Students positively perceived PBL components such as authentic problems, collaboration, and teacher facilitation, though some reported difficulty with limited lectures, autonomy, and workload. CEQ results revealed no significant differences in good teaching, generic skills, and student support, but the GenAI-PBL group reported heavier workloads and lower overall satisfaction compared with the conventional group. Thematic analysis confirmed the value of collaboration, facilitation, and real-world problems, while highlighting challenges in pacing, role specialization, and overreliance on ChatGPT. The study concludes that PBL, with or without GenAI, can foster engagement and reasoning but does not significantly improve statistical thinking scores compared to conventional teaching. Recommendations include embedding mini-activities to strengthen sampling, test justification, and EDA; adopting a blended model that combines GenAI support, PBL collaboration, and conventional structure; refining the TST for validity; extending statistical investigations to broader topics; and exploring structured and ethical GenAI use in future studies.