Аннотация
Generative AI systems are increasingly adopted as on-demand tutors in informatics and introductory programming, yet evidence remains mixed regarding whether performance gains come at the cost of reduced academic independence and heightened over-reliance. Resolving this issue requires moving beyond open-ended AI assistance and toward tutoring systems where pedagogical safeguards are built into the design from the beginning, rather than added later as an afterthought. In this study, we proposed and evaluated an Independence-by-Design pedagogical model for generative AI tutoring that targets improvements in problem-solving quality while preserving learners' academic independence through scaffolded, guardrail-based assistance. The intervention operationalizes tutoring policies such as staged hints, explain-first prompting, and answer-suppression aligned with a programming problem-solving rubric, and it was evaluated in a controlled classroom study comparing a guardrail-based tutor condition with a control condition that used standard instructional resources without AI assistance. Primary outcomes included problem-solving quality, assessed via rubric-based scoring of correctness, reasoning, and code quality, and academic independence, quantified using an Academic Independence Index that triangulates self-regulated learning measures, help-seeking behaviors, and interaction trace features (e.g., hint depth, revision cycles, and direct-answer requests). Effects were estimated using regression models with prior achievement covariates and reported with standardized effect sizes. Results indicated that the guardrail-based generative AI tutor improves problem-solving quality relative to baseline support while maintaining or increasing academic independence indicators, and interaction analyses suggest that moderate hint usage and explain-first behavior are associated with the strongest outcomes. These findings support the responsible deployment of generative AI tutors in informatics education by demonstrating that learning gains can be achieved without amplifying dependency when pedagogical guardrails are embedded in the tutoring design. In practical terms, we give educators a transferable design approach for using generative AI in programming courses. This approach includes staged hints, requiring students to attempt a solution before receiving help, avoiding direct code answers, and prompting students to explain their thinking first. Together, these features help AI support students' reasoning rather than replace it.