With the goal of bridging the gap between group instruction and individualized tutoring, this study explores the use of generative artificial intelligence (AI) in high school programming class, inspired by Bloom’s 2 Sigma Problem. An AI assistant for C++ programming language was set up using ChatGPT’s “MyGPTs” feature, with weekly content limitations to ensure alignment with instructional goals. The AI assistant provided concise guidance on syntax, debugging, and conceptual understanding, while avoiding full solutions. Students used the assistant during and after class. To understand students’ interactions with the AI assistant and learning progress, a five-week study was conducted that included a pretest and a posttest. Students’ chat records, code submissions, and learning outcomes were gathered for both qualitative and quantitative analysis.

Participants in this study were drawn from a group of tenth-grade students who voluntarily enrolled in a computer science project-based elective course. A pretest consisting of three problem-based programming tasks was used to identify suitable participants. Only those who successfully completed at least one of three programming tasks were included in this study. In all 21 students were included and participated in this five-week long study. These 21 students had basic experience using variables and if-else statements to solve simple problems, though they had limited exposure to more structured or multi-step programming tasks. They participated in a four-week course with three hours of instruction per week, focused on using nested loops and arrays to engage in real-world programming tasks. Based on pretest results (completing 1, 2, or 3 tasks), students were categorized into lower-, mid-, and high-performing groups. A posttest was administered in the fifth week with no AI assistance, and all code submissions were evaluated using test cases and the Structure of the Observed Learning Outcome (SOLO) taxonomy.

The results showed that students with lower pretest scores demonstrated the most notable gains in confidence and logical problem-solving skills. Mid-performing students showed steady improvement and used the assistant primarily to reinforce their understanding of core concepts. In contrast, higher-performing students leveraged the AI assistant to extend their learning beyond the curriculum, engaging with more advanced or exploratory tasks. Most students were able to complete 3–5 problem-solving programming tasks each week during class time, without needing to continue the work at home, which is an improvement compared to previous semesters. Students reported that the AI assistant reduced programming frustration and supported task completion, but also noted that its feedback occasionally exceeded their understanding and lacked sensitivity to their learning level, something a human teacher can do better than the AI assistant. This study offers timely insights into how generative AI can emulate aspects of one-on-one tutoring in classroom settings. While students valued the immediacy and clarity of AI support, limitations in adaptivity remain. Future research should explore how such systems can diagnose students’ conceptual development and adjust feedback accordingly, to fully realize the potential of adaptive and personalized instruction.

Keywords: Generative AI, High school programming, Personalized tutoring, Adaptive learning, Computational thinking.