If an Agent Can Write Your Code, What Is Left for You to Learn? Asking, Reviewing, and Owning AI-Assisted Programming

Abstract

When a coding agent can produce a program, its tests, and an explanation, the submission says little about what the student understands. My dissertation investigates how computing students can use AI while retaining the reasoning needed to learn programming. I take part in examining three practices: asking for tutoring that requires students to reason, reviewing generated code against expectations formed before seeing it, and owning the work by explaining AI assistance and their own decisions. In two prompting studies, including a randomized trial with 979 CS1 students, students’ tutoring prompts improved after instruction. In CodeCue, predicting program behaviour before seeing an agent’s changes made students feel better prepared, but did not measurably change how they reviewed. Our study on AI attribution found that what students think should be disclosed differs from what they would disclose themselves. The next phase connects these practices, beginning with a randomized attribution study and extending to repeated tasks with gradually reduced support. I will examine how students learn to direct coding agents, evaluate their work, and retain programming understanding that they can apply independently.