This paper examines whether customizable written exams can effectively assess student understanding of software engineering principles in project-based courses. While repository metrics such as code quantity are easy to capture, learners focus on volume production instead of quality. This is further complicated by the introduction of generative AI, allowing for rapid code production of questionable quality rather than meaningful learning. Written exams provide a way to hold students individually accountable and test their understanding across design, testing, review, and collaboration. We compare two course offerings where student assessment was augmented with exams, uncovering the use of AI-driven superficial coding practices characterized by high volumes of project contributions without a corresponding depth of understanding, deliberate design, or critical review. Although 86% of students reported understanding most AI-generated code, 31% admitted they understood less than they reviewed, pointing to limited critical engagement and cognitive debt. Exams provide a necessary complement to repository metrics, functioning as a validity check that screens for superficial engagement and directs students toward accountable, outcome-aligned learning.

@inproceedings{paperbot12019,
  title = {The Efficacy of Exams in Assessing Student Learning in Capstones},
  author = {Scott Fazackerley and Bowen Hui},
  booktitle = {ICSE},
  year = {2026},
  doi = {10.1145/3786580.3786989}
}