A student submits a program that passes every test. Then the lecturer asks why a particular condition is needed, and the student cannot explain it. That gap matters more than whether the code was written with an AI assistant, copied from a forum, or assembled through trial and error. In each case, a working program tells us something about the program. It tells us less about what the student understands.
Generative AI makes this distinction harder to ignore. It can explain an error message at midnight, suggest a test case, or produce a plausible solution in seconds. Those are genuinely useful capabilities, especially when a student has no one nearby to ask. The educational question is what the student does with that help. Do they become better able to reason through the next problem, or do they become better at obtaining an answer they cannot defend?
What the evidence can tell us
One recent experiment offers a clear warning against treating task scores as a measure of learning. In a 2026 randomized study of 275 introductory programming students at the Technical University of Munich, students worked with either a hint-based AI tutor, unrestricted ChatGPT, or conventional web resources. Both AI groups earned higher scores on the programming exercise and reported less frustration. Neither group showed greater gains in conceptual knowledge or code comprehension on the study’s measures. The hint-based tutor increased intrinsic motivation. These were results from one 90-minute exercise on concurrency, so they should inform course design rather than settle the question for every subject and setting (Bassner et al., 2026).
Another study followed students using generative AI for self-directed programming. It found potential benefits for performance and confidence, alongside weaker long-term knowledge transfer and patterns of overreliance. The stronger learners were more likely to question AI output and work through its meaning. Because this was a quasi-experiment, it is best read as evidence that how students use the tool matters, rather than proof of one universal effect (Li, Liu, & Dong, 2025).
Nor is asking AI for code a skill that novices automatically possess. In a controlled study involving 120 beginning programmers across three institutions, participants struggled to describe problems to a code-generating model and revise their requests when the results were wrong, even with tasks pitched at their level. Good prompting still depends on understanding the problem and judging the result (Nguyen et al., 2024).
Together, these studies support a modest conclusion: AI can make programming work easier, but ease and learning can diverge. That divergence is where teaching has to do its work.
Give students something worth deciding
Human agency in a computing course is practical. It is the ability to frame a problem, choose an approach, notice a doubtful result, test an assumption, and take responsibility for the software one releases. An AI assistant can contribute to each step. The student still needs a reason for accepting its contribution.
Consider a first programming assignment that processes a list of student scores. A student might sketch the inputs, outputs, and edge cases before opening an AI tool. If the code fails, they could ask for a hint about the failing case or an explanation of the error message. Later, they might compare their solution with an AI-generated alternative and test both on empty lists, invalid values, and boundary cases. The important record is the decision: What did I change, why did I change it, and how did I check that it works?
This kind of sequence gives AI a place in the work without hiding the work of learning. It also leaves room for different stages of expertise. Early in a course, an instructor may ask students to trace a loop or write a small function independently. Later, when the target is system design, accessibility, or testing, asking AI to draft routine code may be entirely sensible. The boundary should follow the learning objective and be stated plainly for each assignment.
Assessment should follow the same logic. A finished repository is useful evidence, but it cannot carry the whole judgment. Ask students to explain one design choice, predict the output of an unfamiliar test, diagnose a planted bug, or revise a requirement in a short live demonstration. These checks can be brief and supportive. They reveal understanding that a polished submission alone may conceal. The ACM task force’s 2026 report describes educators moving toward code comprehension, debugging, oral checks, and clearer rules for AI use; it documents emerging practice, not a single proven recipe (Gordon et al., 2026).
Students also need fair conditions for making these choices. A policy that silently assumes everyone can pay for the same tool is a poor policy. So is an assignment that requires students to paste private data, classmates’ work, or confidential project code into an external service. Offer an accessible route to complete the learning task, specify which data may be shared, and make the AI rules visible before students start. UNESCO’s guidance places privacy, equitable access, and human agency within the design of educational AI use, rather than leaving them as afterthoughts (Miao & Holmes, 2023).
AI tools will keep changing, and so will the ways students use them. A durable computing education can make room for those tools while continuing to ask an old, demanding question: Can you explain and stand behind what your software does? When students can answer it, assistance has served its purpose.
References
- Bassner, P., Lenk-Ostendorf, B., Beinstingel, R., Wasner, T., & Krusche, S. (2026). Less stress, better scores, same learning: The dissociation of performance and learning in AI-supported programming education. Computers and Education: Artificial Intelligence, 10, Article 100537.
- Gordon, S., Denny, P., Keuning, H., Kiesler, N., Kumar, A. N., Kumar, V., Leinonen, J., & Prather, J. (2026). ACM Task Force on Generative AI and Programming Assessment: Final report.
- Li, S., Liu, J., & Dong, Q. (2025). Generative artificial intelligence-supported programming education: Effects on learning performance, self-efficacy and processes. Australasian Journal of Educational Technology, 41(3), 1–25.
- Miao, F., & Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO.
- Nguyen, S., Babe, H. M., Zi, Y., Guha, A., Anderson, C. J., & Feldman, M. Q. (2024). How beginning programmers and code LLMs (mis)read each other. Proceedings of the CHI Conference on Human Factors in Computing Systems.