← 返回研究新聞
Editorial cover for a review of generative AI in programming education
綜述證據綜述20252026年6月25日· 8 min

Literature Review on the Integration of Generative AI in Programming Education

Jemimah Nathaniel, Solomon Sunday Oyelere, Jarkko Suhonen, Matti Tedre

International Journal of Artificial Intelligence in Education

500 字摘要

Editorial cover for a review of generative AI in programming education

Nathaniel, Oyelere, Suhonen, and Tedre review a question that is now central to computer science education: how can generative AI tools be integrated into programming education without weakening students' foundational logic, problem solving, and higher-order thinking skills? The paper is useful for AIEDHK because it moves beyond generic enthusiasm for ChatGPT or Copilot. It asks whether the tools are embedded in teaching methods, assessment routines, and learning processes that still require students to understand code rather than only generate it.

The review synthesizes 40 empirical studies using PRISMA 2020 and Kitchenham-style review methods. Its focus is not simply whether GenAI can solve programming tasks. Instead, it examines how studies connect GenAI tools with programming curricula, teaching methods, assessment designs, integration processes, and student skill development. That framing is important because programming education has a long history of intelligent tutoring systems, automated feedback, Parsons problems, code explanation tools, and step-based support. GenAI adds flexibility and natural-language interaction, but it also increases the risk that learners accept generated code without understanding algorithms, syntax, data structures, or debugging logic.

The paper's findings are deliberately implementation-focused. The authors argue that successful integration depends on intentional teaching strategies, thoughtfully designed assessments, and structured integration processes. They also identify barriers: limited accessibility support, insufficient bias mitigation, weak curriculum alignment, and tool selection driven by availability rather than educational fit. These are practical concerns for any school or university considering AI-assisted coding. A tool that improves productivity for experienced developers can still be harmful for novice learners if it bypasses the struggle needed to build mental models.

The review also proposes a GenAI-Ped framework that combines self-regulated learning, universal design principles, prompt-engineering support, and iterative feedback. For AIEDHK, this is the most actionable contribution. It suggests that GenAI in coding courses should be framed as a guided learning partner, not an answer machine. Students need orientation on when to ask for help, how to inspect generated code, how to explain a solution, and how to reflect on what they have learned. Teachers need assessment formats that reveal reasoning, not only final code output. Product teams need interfaces that encourage explanation, comparison, revision, and metacognitive checks.

The paper is especially relevant for Hong Kong because programming education is multilingual, high-stakes, and often linked to future workforce claims. GenAI coding support can make programming more accessible, but only if it is aligned with local curricula, language needs, teacher capacity, and assessment expectations. AIEDHK can use this review to evaluate AI coding tutors, coding assistants, and student copilots through a clear test: does the system help learners develop durable programming understanding, or does it mainly make correct-looking code easier to obtain?

相關論文

A Black male programming student and an East Asian female peer trace a loop on a small whiteboard while revising a printed multiple-choice question in a university maker room
期刊論文2026
期刊論文 218

Blending guidance with reflection prompted more revisions than reflective AI feedback alone

Omar Alsaiari, Nilufar Baghaei, Jason M. Lodge, Omid Noroozi, Dragan Gašević, Marie Boden, Hassan Khosravi

Computers and Education: Artificial Intelligence

A 12-week study followed 329 students receiving directive, reflective, or hybrid AI feedback while creating learning resources. Hybrid feedback had the highest observed content-revision rate, and both hybrid and directive feedback exceeded the reflective condition. Hybrid and directive conditions did not differ significantly from each other, and final resource quality showed no significant group difference. The study makes revision behavior visible while leaving deeper understanding and transfer unresolved.

feedback designstudent-generated resourcesmetacognition
閱讀 500 字摘要 →
A South Asian programming lecturer and a white student test a small equipment-booking application beside handwritten bug-reproduction steps in a warm computer classroom
政策 / 倫理2026年10月16日
政策 / 倫理 219

Product news: Claude Code Review becomes generally available, making review decisions a teaching opportunity

Anthropic

AI Product and Learning Report

Anthropic made Claude Code Review generally available for Team and Enterprise organizations on October 15. This October 16 report examines how programming courses might use automated findings as material for explanation and testing. A proposed review exercise asks students to reproduce a reported defect, defend or reject the diagnosis, and verify a correction. Faster product workflows do not establish stronger programming skills, and a completed automated review does not remove responsibility for the final code.

product newsClaude Codecode review
閱讀 500 字摘要 →
A programming student traces an algorithm and test cases on paper while a tutor points to a test report beside code on a laptop.
期刊論文2025
期刊論文 134

Students valued AI coding help while exam evidence exposed difficulties in correcting generated code

Dan Kohen-Vacs, Maya Usher, Marc Jansen

International Journal of Artificial Intelligence in Education

Two studies examined student attitudes to generative AI and performance on code-correction tasks. A survey found favorable perceptions, while exam analyses found lower performance on AI-generated debugging tasks in several cohorts. The studies involved different students and compared debugging with code construction, so they do not establish that positive attitudes caused weak verification or that AI origin alone made the tasks harder. They identify a specific skill worth teaching and assessing.

programming educationdebuggingassessment
閱讀 500 字摘要 →