Inquiry-Based Learning in STEM Education: The Impact of Generative AI-Based Chatbots on Primary School Students' Problem Posing Ability in Science

Synthesis: This quasi-experimental study (N = 97 Chinese third-graders, split by class) compared GenAI-based chatbots against search engines for fostering science problem posing in Inquiry-Based Learning. Chatbots significantly improved problem quality (t = 2.47, p = 0.015) and overall problem posing ability (t = 3.07, p = 0.003), produced a more integrated epistemic network structure (ENA), and reduced cognitive load while raising technology acceptance.

Key Findings

1. Chatbots beat search engines for problem posing. The experimental chatbot group outperformed the search-engine control on problem quality (t = 2.47, p = 0.015) and total problem posing ability (t = 3.07, p = 0.003), with no significant difference in the number or category of posed questions.
2. More integrated cognitive networks. Epistemic network analysis (ENA) showed the chatbot group developed a broad, high-density, integrated network (e.g., applying → scientific-question contexts), whereas the control group exhibited local clustering.
3. Lower cognitive load, higher acceptance. The chatbot group reported lower cognitive load and higher technology acceptance than the search-engine group.
4. Three rounds of progressive inquiry. Following three rounds of progressive problem-based inquiry learning, students' problem posing ability increased significantly in both groups (number and quality dimensions), with chatbots amplifying the effect.

What this means for practice

• Instructors. Use the chatbot to scaffold the questioning phase of inquiry, not to supply answers: the chatbot group improved the quality of posed problems and total problem posing ability while reporting lower cognitive load.
• Instructors. Structure inquiry in progressive rounds: both groups improved problem posing across three rounds of problem-based inquiry and the chatbot amplified the effect, so pair sustained dialogue with a repeated questioning cycle rather than a one-off task.
• Designers. Choose conversational dialogue over keyword retrieval when the goal is questioning: the chatbot's "questioning–answering–follow-up questioning" cycle is what distinguished it from the search-engine group's one-way retrieval.
• Designers. Keep the search engine as a legitimate alternative for simple information lookup: this study found no significant difference between the groups in the number or category of posed questions, only in their quality.

Limitations

• N = 97 third-grade students from a single school, split by class rather than randomized, so class-level differences and the specific classroom design are confounded with condition.
• The intervention was brief and measured only problem posing within one science topic; the authors note they did not track conceptual deepening or experimental inquiry, leaving effects on the full inquiry chain untested.
• Both conditions used specific platforms, and the authors state that differences in platform functions, prompt design, and teacher support may influence how the tools perform, limiting transfer to other configurations.
• The sample was third-graders in one school, so applicability to other age groups, school contexts, and longer-term instruction still needs validation.

Citation

Dai, Z., Huang, F., Xiong, J., Yang, Y., Zhang, Q., & Peng, X. (2026). Inquiry-based learning in STEM education: the impact of generative AI-based chatbots on primary school students' problem posing ability in science. International Journal of STEM Education, 13, 44.