Guided ChatGPT inquiry accompanied stronger thermodynamics scores, with teacher support built into the intervention

An AI explanation of heat, work or entropy can sound clear while leaving a student's underlying misconception intact. Tarik El Fathi and colleagues investigated a teaching approach that required learners to question and evaluate those explanations rather than simply collect answers. Their thermodynamics study combined ChatGPT with a constructivist inquiry-based prompting framework. The educational object was a guided cycle of asking, checking and revising, with the instructor still responsible for clarifying concepts and addressing difficulties.

The quasi-experiment involved 120 first-year engineering students at Ibn Tofail University in Morocco during the autumn semester of 2023. Sixty students followed conventional instruction and sixty used the AI-supported approach. Groups were formed with comparable academic profiles and taught by the same instructor; the study was not an individually randomized trial. Conceptual understanding was assessed before and after teaching with a French adaptation of the Survey of Thermodynamic Processes and First and Second Laws. Weekly questionnaires also tracked experiences and acceptance of ChatGPT.

Before using the system, the intervention group completed three hours of training in constructing prompts, checking responses and refining questions. During the semester, both groups attended two-hour weekly tutorials. The AI-supported sessions included prompt refinement, evaluation of generated explanations, collaborative discussion and instructor feedback. The comparison group used worked examples, guided exercises and conventional problem solving. The matched tutorial schedule helps describe the comparison, but AI access was bundled with a distinctive teaching approach and preparation rather than varied as an isolated ingredient.

The intervention group achieved a higher conceptual post-test mean of 22.53, compared with 18.60 in the comparison group. Both groups improved, while the AI-supported group showed the larger gain. The authors reported clearer progress on some conceptual topics and persistent difficulty with areas involving quantitative reasoning, including engine efficiency and reversible processes. Exploratory analyses of item patterns suggested different conceptual structures across groups. Those statistical patterns may guide further investigation, but they do not directly show the structure of an individual learner's understanding.

The findings should remain specific to the instructional package and context. Convenience sampling at one institution, nonrandom allocation and a single semester limit causal attribution and generalization. The intervention used the available free ChatGPT service, whose behavior can change over time. Participants also had reliable computers and internet access during tutorials, conditions that may not hold elsewhere. Positive acceptance ratings are useful implementation evidence, but they cannot establish conceptual mastery, independent problem-solving skill or retention after the assistant is no longer available.

For AIEDHK, the practical next step would be a small science tutorial that preserves a learner's initial explanation, the AI response and the reason for a revision. Teachers could contrast conceptual questions with numerical problems and require students to check assumptions, units and physical plausibility. A delayed unaided test would help assess whether the learning persists. These are proposed extensions beyond the study. The useful lesson is to evaluate guided inquiry as a whole, with teacher review and independent performance checks, rather than attribute every observed improvement to access to a conversational model.