As generative AI reshapes how programmers learn and work, a new study from U.S. researchers offers crucial insight for computer science educators. The research, to be presented at the Koli Calling International Conference on Computing Education Research in Finland, explores how undergraduate students develop and recalibrate trust in AI-powered programming tools such as GitHub Copilot and ChatGPT.
Short-Term Trust, Long-Term Realization
Over several weeks, 71 computer science juniors and seniors participated in experiments using GitHub Copilot to build new functionalities within open-source codebases. Initially, more than half reported increased trust in the tool after a hands-on session demonstrating its capabilities. But after completing a 10-day programming project, trust levels became more mixed: 39% said their confidence in Copilot grew further, while 37% said it declined as they encountered its limits in complex tasks.
“Using Copilot to its full extent requires a competent programmer who can complete some tasks manually,” said Dr. Gerald Soosairaj, associate teaching professor at UC San Diego and senior author of the paper. The finding underscores that while AI assistants accelerate coding, they don’t replace the need for human comprehension and debugging skills.
Balancing Learning and Assistance
The researchers warn of a dual challenge in AI-assisted programming education. Students who depend too heavily on chatbots risk bypassing critical learning processes, while those who avoid them entirely may fail to acquire essential skills for modern software development. The goal, they argue, is calibrated trust knowing when and how to rely on AI output, and when to question it.
Guidance for Computer Science Educators
The study offers several actionable recommendations for educators adapting curricula to the AI era:
-
Design Mixed-Difficulty Assignments – Give students coding tasks of varying complexity to help them gauge AI reliability across contexts.
-
Maintain Manual Proficiency – Ensure students can read, debug, and test code without AI support.
-
Teach AI Mechanics – Explain how language models generate and refine output to foster realistic expectations.
-
Highlight Key Features – Train students to use AI assistant functions effectively, including commands like “/explain,” “/fix,” and “/docs.”
-
Frame AI Discussions Carefully – Instructor tone and framing shape how students perceive and trust AI tools.
The team plans to expand the study to 200 students in the coming academic term to deepen insights into how trust evolves over time.
The Bigger Picture: Redefining Competence in the AI Era
The study’s implications extend beyond classrooms. As generative AI becomes standard in professional development environments, tomorrow’s programmers must pair technical literacy with AI fluency — understanding both the potential and the pitfalls of machine-generated code.
For educators, this means rethinking what “learning to program” means in a world where AI can code but not understand. The focus shifts from syntax mastery to critical evaluation, debugging insight, and responsible AI collaboration.
Adapted from University of California San Diego.