AI-Powered Course Engines Are Rewriting How Students Learn to Program

Artificial intelligence is no longer a distant promise in higher education; it is being wired directly into the classroom, and one of the most demanding proving grounds is the introductory programming course. A new study published in Frontiers of Digital Education by Xiaohong Su, Xiaofei Xu, and Tiantian Wang of the Faculty of Computing at Harbin Institute of Technology describes a systematic attempt to rebuild programming instruction around large language models, moving beyond the familiar pattern of bolting a chatbot onto an existing learning management system. Their work arrives at a moment when China’s Education Informatization 2.0 Action Plan has made the fusion of AI and teaching a national priority, and when education ministries worldwide are scrambling to understand what generative models mean for the way knowledge is transmitted, practiced, and assessed.

The researchers begin from a sober diagnosis of the problems that have plagued programming education for decades. Introductory courses typically enroll hundreds of students with wildly different backgrounds, yet instruction is delivered in a single uniform stream. Teaching assistants are stretched thin, feedback on code arrives days after the moment of confusion, and students who fall behind early often never recover. These are precisely the failure modes that large language models, with their ability to generate explanations, examples, and code critiques on demand, appear positioned to address. The team’s response was not simply to hand students access to a general-purpose chatbot, but to engineer a dedicated course engine powered by large language models that anchors the AI’s outputs in the specific content, pacing, and pedagogy of an introductory programming curriculum.

At the technical heart of the study is the construction of this course engine. Rather than treating the language model as an oracle, the researchers designed it as the generative core of an intelligent course platform: the engine produces structured teaching scenes, generates practice materials aligned with course objectives, and adapts explanations to individual learners’ demonstrated gaps. The engine drives what the authors call innovation in teaching scenes, meaning that the sequence of lecture, demonstration, exercise, and review is no longer fixed in advance but can be reconfigured dynamically as the system observes how students actually perform. This represents a shift from static courseware, where every student traverses the same path, toward a generative architecture in which the course itself is assembled and reassembled by AI in response to learning data.

Building on the engine, the team designed AI-enhanced teaching and learning environments organized around two complementary agents: AI teaching assistants for instructors and AI learning companions for students. For teachers, the assistants provide scalable, differentiated support, automating the repetitive labor of answering routine questions, generating variants of exercises, and monitoring class-wide progress so that scarce human attention can be directed at the students who need it most. For students, the companions deliver one-on-one, adaptive, and customized learning experiences, offering hints rather than answers, diagnosing misconceptions in code, and adjusting difficulty in real time. The design philosophy is explicitly one of augmentation: the AI absorbs the volume of individualized interaction that no human teaching team can sustain, while preserving the human instructor’s role in motivation, judgment, and mentorship.

Perhaps the most ambitious element of the work is the integrated learning support system the authors propose, which stitches together six stages of the educational experience: courses, training, competitions, testing, evaluation, and certification. The goal is a smart teaching ecosystem in which knowledge services, personalized learning, and instructional support operate continuously across what the researchers describe as all elements and all time periods of the teaching process. In practice, this means that a concept introduced in a lecture flows into AI-generated training exercises, surfaces again in competitive programming challenges, is measured by automated testing, feeds an evaluation profile of each learner, and ultimately connects to certification of demonstrated competence. The pipeline is designed so that no stage is an isolated event; data from each informs the next, closing the loop between instruction and assessment.

A concrete embodiment of this vision is the development of intelligent and interactive virtual massive open online courses, which the authors abbreviate as IMOOCs, for the C programming language. Traditional MOOCs revolutionized access to elite instruction but suffered from notoriously low completion rates, largely because they offered broadcast-style videos with minimal interaction. The IMOOC concept reimagines the format: virtual and interactive elements are embedded directly into the course, allowing learners to converse with the material, receive immediate feedback on code, and progress along individualized paths. The researchers also explored a new hybrid teaching model built on the IMOOC that integrates virtual and real elements and promotes cross-domain collaboration, blending the scalability of online delivery with the engagement of in-person, project-based classroom work.

The study situates these innovations within a broader global and policy context. The authors analyze worldwide trends in AI-powered teaching and learning, citing the growing body of computing-education research on generative AI, including work published in Communications of the ACM examining how tools like code-generating assistants change what students must actually learn when machines can write routine programs. In China, the Ministry of Education has released successive batches of typical application scenarios for AI in higher education in 2024, and the State Council in 2025 issued opinions on deepening the implementation of the Artificial Intelligence Plus initiative, alongside a white paper on smart education. The Harbin team’s programming-course experiment can be read as a ground-level implementation of these top-level strategies, testing in a real curriculum what policy documents describe in the abstract.

Notably, the authors do not present AI as an unalloyed good. The study explicitly discusses the potential risks of overreliance on AI tools, a concern that has intensified across computing education as students increasingly submit machine-generated code they cannot explain. If a learning companion always supplies the next step, the argument goes, students may never develop the struggle-based reasoning that programming instruction is supposed to cultivate. The researchers outline strategies to address these risks, positioning the AI as a scaffold that must be deliberately tapered rather than a crutch that quietly replaces cognitive effort. This candor distinguishes the work from more breathless accounts of educational AI and acknowledges a tension that every institution deploying these tools will have to manage: the same model that can personalize feedback can also short-circuit learning if its outputs are consumed passively.

The implications extend well beyond one C programming course. If course engines built on large language models can reliably generate intelligent, adaptive instruction, the marginal cost of high-quality individualized teaching approaches zero, which has profound consequences for institutions that have rationed personal attention by enrollment size. The integrated course-training-competition-testing-evaluation pipeline also anticipates a shift in how competence is certified, moving from snapshot examinations toward continuous, evidence-based assessment accumulated across an entire learning journey. For educators, the message is that the unit of design is no longer the lecture or the textbook but the ecosystem: the interlocking set of AI services, human roles, and data flows that together constitute a modern course.

Challenges remain substantial. The authors point toward future trends and unresolved difficulties in the AI-plus-higher-education landscape, and their own account implies open questions about cost, model reliability, academic integrity, and the training teachers will need to orchestrate these systems. Yet the study’s central claim is measured and forward-looking: AI will unlock new possibilities for reshaping how higher education is delivered and experienced. What the Harbin Institute of Technology team has demonstrated is that this reshaping need not be a vague aspiration. With a course engine, paired AI assistants and companions, an integrated six-stage support system, and an interactive virtual MOOC, the components of an AI-empowered classroom already exist and have been exercised in practice. The remaining task, the researchers suggest, is to scale them thoughtfully, guarding against dependence while extending to every student the kind of individualized attention that education has always promised but rarely delivered.

Subject of Research: Application of large language model-powered course engines and AI assistants to programming education in higher education

Article Title: Innovation of Teaching and Learning Scenes and Models Empowered by Artificial Intelligence: Practice and Experience of AI-Powered Programming Courses

Article References: Su, X., Xu, X., & Wang, T. (2026). Innovation of Teaching and Learning Scenes and Models Empowered by Artificial Intelligence: Practice and Experience of AI-Powered Programming Courses. Frontiers of Digital Education, 3(1), Article 4. https://doi.org/10.1007/s44366-026-0078-y

Image Credits: AI Generated

DOI: 10.1007/s44366-026-0078-y

Keywords: artificial intelligence, higher education, programming education, large language models, smart education, IMOOC, AI teaching assistants, personalized learning, course engine, educational technology, C programming, teaching reform