Programming students shifted their AI habits with guidance, but better interaction was not a randomized learning effect

Giving students access to ChatGPT does not determine how they use it. Hacer Güner and Erkan Er examined that problem inside an introductory Python course, following how learners asked for code, explanations, debugging, and revision. Their main contribution is a description of changing interaction habits tied to classroom activities. The results support paying attention to the work students do before and after an AI response, while leaving the causal effect of individual support strategies unresolved.

The course ran for 14 weeks across three sections taught by one instructor. Of 175 enrolled students, 158 participated in all research activities, although the usable samples varied between specific analyses. The researchers collected prior-knowledge scores, voluntarily submitted ChatGPT conversations, and quizzes completed without outside help. Across the sessions, they coded 2,388 prompts. This combination provided a closer view of learning processes than satisfaction ratings or submitted code alone would have offered.

Three activities introduced different forms of support in a fixed sequence. Students first tackled a programming task with unrestricted ChatGPT access and no guidance. A later session began with a 20-minute demonstration of practices such as dividing a task, requesting hints, checking generated code, and asking for explanations. The final activity supplied example prompts within a lab guide. The programming topics also changed, moving from variables and functions through loops and conditions to lists, strings, and file handling.

Five interaction profiles emerged, ranging from asking AI to generate a solution to writing independently or using AI to improve an existing attempt. Profile distributions shifted significantly across sessions. By the third session, the collaborative profile was the most common. In the earlier sessions, students who first produced their own code and then requested refinement generally performed better on particular quiz comparisons than students relying on generated solutions. The authors also found that prior programming knowledge differed between some profiles.

Those distinctions limit what the findings establish. The study did not randomly allocate comparable students to alternative prompting lessons, and the same activities were not repeated under interchangeable conditions. Increasing familiarity, changing task difficulty, and accumulated programming knowledge accompanied the interventions. Profile membership was an observed behavior, not an assigned treatment. Better quiz performance among code refiners therefore cannot by itself show that instructing any learner to adopt that profile will produce the same advantage.

The authors further acknowledge incomplete conversation records, possible unreported use of other resources, and a sample from one university with similar backgrounds. Their data capture selected classroom episodes rather than a long-term transfer test. These boundaries matter when interpreting the work as evidence for AI literacy: learning to ask more elaborate questions is useful only if students also understand, test, and eventually reproduce the underlying programming decisions independently.

For AIEDHK, a reasonable classroom application is to make the sequence of student work visible. Ask learners to retain their initial attempt, identify the precise point where they requested assistance, and explain why they accepted or rejected a suggestion. Teachers could then pair that process record with a short unaided exercise. This is an instructional implication of the study, not an intervention it directly validated, and it keeps the focus on developing programming judgment rather than simply producing functioning code.