This exploratory bounded case followed three pre-service science teachers building a Kundt's tube through four Do-It-with-AI phases: conceptual exploration, design development, troubleshooting and extension. Scopus AI supported literature exploration, while ChatGPT-4/5 proposed design alternatives, failure hypotheses and applications. Participants modified, built, tested and decided for themselves. After an annular-gap adjustment, discrepancy fell to about 1.2%; the final apparatus measured sound speed at 346.6 m/s, about 0.1% from the theoretical ambient-temperature value. The case demonstrates a feasible AI-supported workflow, but without a comparison group or pre–post measures it cannot establish that AI improved learning or epistemic agency.
Key findings
- The four-phase workflow separates AI roles in concept search, design alternatives, troubleshooting hypotheses and application ideas while retaining human adaptation and physical verification.
- An annular-gap modification reduced discrepancy to about 1.2%, illustrating iteration between a hypothesis and prototype testing.
- The final configuration measured sound speed at 346.6 m/s, approximately 0.1% from the theoretical value at ambient temperature under the case conditions.
- Evidence of epistemic agency was strongest during troubleshooting, where participants selected and tested suggestions; the extension phase was more retrieval-oriented.
- The four phases should not be read as a developmental ladder, and device accuracy cannot disentangle contributions from AI, instruction, collaboration and ordinary trial-and-error.
Why this matters globally
The framework addresses a global question: where should real-world verification sit in generative-AI laboratory learning? Its strength is treating AI proposals as provisional hypotheses and physical evidence as the arbiter, a potentially useful model for learner accountability in STEM. Scaling requires comparison with non-AI instruction and measures of understanding, transfer, verification skill, dependency and cognitive load—not only final device accuracy.
Thai researcher contribution
Researchers from Walailak University and King Mongkut's Institute of Technology Ladkrabang combined physics, teacher education and music/engineering in a concrete case. Beyond proposing a framework, the Thai team exposes within-case variation in learner roles across task types, providing an appropriately cautious basis for the next generation of AI-education trials.
Limitations to consider
Only three participants, one activity and one apparatus were studied, making this feasibility/process evidence rather than an effectiveness estimate. There was no pre–post measure, comparison group, complete discourse analysis, or independent assessment of conceptual learning, AI literacy, epistemic agency or cognitive load. Device accuracy is a joint outcome of several influences and does not prove individual learning. Reported agency is activity-level distribution, not durable within-person change, and transfer to other subjects remains a proposition.