← Zurück zu Forschungsnachrichten
Editorial cover for a randomized comparison of AI support in university programming education
ZeitschriftenartikelPeer-reviewed study202619. Juli 2026· 3 min

Less stress, better scores, same learning: The dissociation of performance and learning in AI-supported programming education

Patrick Bassner, Ben Lenk-Ostendorf, Ramona Beinstingel, Tobias Wasner, Stephan Krusche

Computers and Education: Artificial Intelligence

500-Wörter-Zusammenfassung

Editorial cover for a randomized comparison of AI support in university programming education

Bassner, Lenk-Ostendorf, Beinstingel, Wasner, and Krusche examine a central problem in AI-supported education: does better performance while using AI mean that students have learned more? Their open-access study reports a three-arm randomized controlled trial in an introductory programming course at the Technical University of Munich. The design compares two forms of AI assistance with a no-AI control and measures task success, conceptual learning, cognitive load, frustration, and motivation.

The 275 participants completed a 90-minute exercise on concurrency in which they implemented a parallel sum using threading. One group used Iris, a scaffolded tutor designed to provide calibrated hints while withholding complete solutions. A second group used unrestricted ChatGPT, which could return full solutions. The control group used conventional web resources without an AI assistant. This contrast is valuable because it separates the presence of generative AI from the instructional design wrapped around it.

The researchers measured performance through automated test coverage for the programming exercise. Learning was assessed through pre- and post-knowledge tests and a code-comprehension task. Validated scales captured intrinsic, germane, and extraneous cognitive load, frustration, and intrinsic motivation. These separate measures let the study test whether students who complete more of the exercise also show greater conceptual gains, rather than treating a working submission as sufficient evidence of learning.

Both AI-supported groups achieved substantially higher exercise scores than the control group. The score distributions differed: ChatGPT users clustered toward high task scores, control participants clustered lower, and Iris users were spread more widely across the performance range. Students in both AI conditions also reported less frustration and lower extraneous and germane cognitive load than the control group, while intrinsic cognitive load did not differ. Participants rated ChatGPT as easier to use and more helpful.

The learning results tell a more cautious story. Neither AI condition produced larger pre-to-post knowledge gains or a code-comprehension advantage over conventional resources. In this short programming exercise, generative AI therefore functioned mainly as a performance aid. The scaffolded Iris tutor did show one distinctive benefit: it increased intrinsic motivation, while unrestricted ChatGPT did not. The authors describe the attraction of easy, helpful assistance without additional learning as a comfort trap, where learners' preferences and immediate success can diverge from pedagogical effectiveness.

The study should not be generalized beyond its boundaries. It examines one 90-minute concurrency task in one introductory university programming context, and its learning measures focus on knowledge change and code comprehension around that activity. Longer courses, different domains, repeated tutor use, or designs that require explanation and transfer may produce different outcomes. The open dataset and analysis materials strengthen transparency and make replication or secondary analysis possible, but broader evidence is still needed.

For AIEDHK, the practical implication is to separate assistance metrics from learning metrics. A school or university pilot should not infer understanding from completion rate, code quality, reduced frustration, or positive user ratings alone. AI-supported tasks can require students to predict program behavior, explain a generated solution, debug a novel case, compare alternatives, and complete an independent transfer assessment. Scaffolded hints may better preserve motivation and productive activity than unrestricted answer delivery, but their value still needs to be demonstrated through learning evidence. The paper's most important contribution is a clear warning: when AI makes work easier, assessment must work harder to reveal what the learner can actually understand and do.

Verwandte Beiträge

Illustrative photograph of a teacher-education group discussing a short fictional tutoring exchange, with two alternative response cards on a seminar table.
Tool / Datensatz2026
Tool / Datensatz 249

Preprint: TutorCaseBench separates recognizing a strategy from responding to a learner

Yeon Su Park, Chanyoung Park, Tae Soo Kim, Seungju Kim, Hyunwoo Kim, Daho Jung, Jinsung Yoon, Juho Kim

arXiv preprint

A benchmark builds simulated tutoring cases from qualitative education studies. Models were more successful at choosing a proposed strategy than at generating a response adapted to the particular situation.

preprintAI tutor evaluationcase-based assessment
500-Wörter-Zusammenfassung lesen →
Illustrative photograph of a teacher and two learners exploring a layered cutaway of the Earth on a classroom display, with a matching student sketch.
Politik / Ethik29. Okt. 2026
Politik / Ethik 248

Product news: Gemini immersive explorations turn visual topics into layered learning journeys

Google

AI Product and Learning Report

Google’s October 28 release introduces interactive topic explorations through the Gemini app and student hub. Teachers can connect the visual layers to predictions and explanations while checking language availability and learning evidence.

product newsGemini for Educationimmersive explorations
500-Wörter-Zusammenfassung lesen →