<![CDATA[The use of large language models (LLMs) in programming education is growing rapidly, yet evidence on its benefits, risks, and best practices has not been mapped in an integrated way. This study aimed to map the characteristics of empirical studies, synthesize benefits and risks including plagiarism and overreliance, and identify evidence-supported instructional practices. This systematic literature review followed PRISMA 2020. Searches were run through Consensus on 9 October 2026 using 14 queries, yielding 140 records. Screening, extraction, and six-item quality assessment were performed by a single AI-assisted reviewer based on abstracts. A total of 68 studies published between 2023 and 2026 were included and synthesized thematically. The most frequently reported benefit was on-demand feedback and debugging help (30 studies), followed by motivation and self-efficacy (9). Of 21 studies reporting learning outcomes, 10 were positive, 5 mixed, 5 null, and 1 negative. Main risks were overreliance (22 studies), unreliable or solution-leaking output (20), and academic integrity concerns (9). Promising practices include pedagogical scaffolding, prompt literacy training, self-regulation support, process-based assessment, human oversight, and explicit AI-use policies. Findings are preliminary because the search was limited to one aggregator and screening relied on abstracts.]]>
Copyrights © 2026
