Question and Answer Feedback for AI Safety Class

Free-text feedbackTen respondents answered Q5 (“Feel free to write a review about this class (e.g., What did you learn?” “If you met your goal, why? If not, why not?” “What should you study next?”) and ten answered Q6 (“What did you think about this course? Please share your opinions about good and aspects to be improved (e.g., course schedule, materials & instruments, responses towards students’ inquiries, level of exams).”). Because most comments were brief (1–3 sentences), we coded themes as present/absent rather than graded intensity. Three gains surfaced repeatedly: (1) breadth and foundations (“I got a basic understanding of ML and AI fundamentals”), (2) alignment/safety awareness (“learned about GenAI risks and alignment, especially agent safety”), and (3) confidence to go deeper (several students planned follow-up study with code). Suggestions for improvement coalesced around four ideas: (1) a desire for some hands-on artifacts or small code-adjacent demonstrations (“more technical with actual coding”), (2) guest-lecture calibration (“too many guest lectures—prefer more time for lecture or discussion” or “fewer but deeper talks”), (3) deeper, more specific cases tied to current topics, and (4) more structured peer exchange beyond reading and short presentations. These comments help explain the small tail of lower satisfaction and proactiveness in Figure 3. We include the full anonymized responses and the coding key in the Technical Appendix (supplementary), Table 3, together with brief excerpts that illustrate each theme. To preserve anonymity and space in the main paper, we avoid extended quotations here and report only the dominant patterns that informed our course revisions.