Introduction
Nearly every undergraduate uses generative AI now. Universities have mostly responded by trying to catch them: detection software, tightened integrity policies, and, as two students in my study discovered, even hidden instructions buried in assignment briefs designed to trap anyone who pastes the task into a chatbot.
One student I interviewed wrote his essay himself. A classmate generated his, then ran it through a "humanizer" tool, and scored higher. Nothing caught it, and my interviewee confessed he felt betrayed. That episode is easy to read as a policy failure; however, I argue that it is a design failure. Moreover, I noticed that in all the discussion about how to redesign assessment, almost nobody was asking the people actually sitting these assessments - students.
Methodology
Over the summer I interviewed twelve HKU undergraduates from ten different disciplines - engineering, history, psychology, journalism, chemistry, computer science, philosophy and others. I asked them how they use AI, what kinds of assignments they had had, which taught them the most, which were easiest to fake, and what they would change if they were dean for a day.
Findings
Three kinds of assessment resisted AI. Ones that remove access to it, like closed-book exams. Ones that require you to perform live, like presentations and podcasts. And ones built on the student's own experience - internships, or a psychology assignment where students ran a two-week behaviour-change experiment on themselves and then explained it using course theory. A model can write about a self-experiment, but it cannot fake having done one.
The more interesting finding was about why students outsource their coursework. It usually was not because a task was boring or badly designed. Students described handing work to AI in courses they genuinely enjoyed, when too many assignments were stacked into one semester, or when a task demanded full effort for a small share of the grade. When I asked what they would change, not one of them suggested better detection, as they generally talked about "more rational" workload and better weighting.
Conclusion
Assessment security is usually treated as an enforcement problem. My interviews suggest it is at least partly a design issue and that students have a clearer view of it than they are usually given credit for.