Human–genai Interaction Modes in Programming Education

Generative AI (GenAI) tools are becoming part of everyday learning in higher education, especially in programming courses where students can ask for explanations, debugging help, or complete code. This creates an important educational question: not only whether students use AI, but how they interact with it. This study compares three Human–GenAI interaction modes in an introductory object-oriented programming course: No AI, AI Tutor, and Direct AI. Thirty students completed three C# programming tasks under a within-subject classroom design with a balanced rotation of task order. The study examined completion time, code quality, measured understanding, perceived difficulty, and students’ subjective evaluations. The Direct AI task–mode combination was associated with the strongest product-oriented outcomes: shorter completion time (Mdn = 10 min), higher code scores (M = 9.20/10), and the highest completion rate (96.7%). AI Tutor and No AI showed longer completion times and lower code scores. However, measured understanding varied little across modes (M = 3.57–3.70/4), with all medians at the maximum score, indicating a ceiling effect. The raw-score product–understanding gap was largest under Direct AI (M = 5.50), reflecting a stronger increase in code scores than in measured understanding. Student perceptions added a different perspective: 70.0% selected AI Tutor as the most effective mode for learning, while 90.0% selected Direct AI as the fastest. The findings suggest that unrestricted AI support can optimise task completion and product quality, but these gains should not be read as equivalent to learning gains. Guided AI Tutor interaction may be educationally valuable because it is designed to preserve more of the student’s reasoning process. Because each mode was paired with a specific task, the results should be interpreted as task–mode combination differences rather than isolated causal effects of interaction mode. The study contributes empirical evidence for designing GenAI-supported programming activities and assessment practices that distinguish product-oriented performance from learning-oriented outcomes.