"Is this a trap?": Student teachers' perceptions and adoption of GenAI in assessments in three teacher education courses

Synthesis: Zou, Hounsell, Huijser, and Tse (2026) report an explanatory sequential mixed-methods study of how student teachers interacted with generative AI (GenAI) in three teacher education courses at a research-intensive Hong Kong university, where instructors explicitly permitted its use in open-ended, reflective assessments. Drawing on an anonymous survey (N = 85), 11 Cantonese-language interviews, course materials, and 158 assignment declarations, the study found that only 37.6% of students chose to use GenAI at all — far below the 79–83% adoption reported in comparable UK and Australian studies — and that those who did used it mainly for proofreading and clarity checking rather than text generation. Through a socio-technical lens the authors locate the explanation not in technology but in sociocultural factors: assessment design, the perceived learning value of the task, self-confidence, and fear of being wrongly accused of plagiarism. Nine of 11 interviewees described the permissive course policy as a potential "trap", a suspicion seeded by inconsistent GenAI rules in other courses the students took. The authors conclude that consistent, program-level assessment policy, exemplars, and feedback matter more than brief course-level instructions.

Key Findings

• Most students declined GenAI even when it was allowed: Only 37.6% (32 of 85 surveyed students) interacted with GenAI in their assessments, while 62.4% (53) chose not to — markedly lower than the 79% reported at King's College London (Gonsalves, 2024) and 83% in an Australian study (Matthews et al., 2024), challenging the assumption that students exploit GenAI when invigilation is absent.
• Adoption diverged sharply by course and level: The undergraduate Course A saw 14 of 23 students adopt GenAI, whereas the two postgraduate diploma courses saw majorities declining (Course B: 29 of 39 not adopting; Course C: 15 of 23).
• Proofreading dominated, text generation was rare: Among the 32 adopters, proofreading was the most cited purpose (14, 43.8%), followed by checking whether their writing was clear (11, 34.4%); only 6 students used GenAI to generate texts, and every adopter reported some benefit (8 said it helped a great deal, 14 to some extent, 7 a little).
• Self-declaration under-reported use: Of 158 enrolled students, declarations submitted with assignments diverged from survey responses in every course — Course A 8 declared vs 14 surveyed, Course B 6 vs 10, Course C 12 vs 8 — suggesting students feared declared GenAI use would affect their grades while anonymous survey answers would not.
• Working alone and plagiarism fear drove non-adoption: Non-adopters most often said they preferred to work by themselves (41, 77.4%) or were concerned about being accused of plagiarism (22, 41.5%); only 7 (13.2%) cited lacking the knowledge or skills to use GenAI.
• The permissive policy was read as a "trap": Nine of 11 interviewees worried about plagiarism accusations, and the word "trap" recurred repeatedly — students suspected the allowance was a lure to catch those who could not resist, a suspicion shaped by inconsistent GenAI policies across other courses and by a one-month institutional ban in 2023.
• Engagement was high and GenAI-neutral: Assessment engagement averaged 4.21 out of 5 (SD = 0.46), with no significant difference between adopters and non-adopters (U = 781.5, r = 0.07, p = 0.536), reinforcing that engagement hinged on assessment design, not tool use.
• Student teachers saw little relevance to their own classrooms: Secondary-sector participants assumed closed-book school examinations would expose any GenAI misuse, and kindergarten teachers saw GenAI as useful for administration rather than teaching — with almost no transfer of their own assessment experiences into their future practice.

Study Design & Method

• Explanatory sequential mixed-methods design (Creswell & Creswell, 2022) with a survey followed by semi-structured interviews to explain the quantitative patterns.
• Three purposively sampled teacher education courses (one undergraduate, two postgraduate diploma) in pre- and in-service programs at a research-intensive Hong Kong university, delivered September–November 2024; all assessments were open-ended and reflective except a closed-book invigilated quiz in Course B.
• Survey of N = 85 of 158 enrolled students (response rates: 100% Course A, 60% Course B, 32.9% Course C), covering assessment engagement (Evans & Zhu's scale, adapted), GenAI adoption in the highest-weighted assessment, and demographics, plus three open-ended questions.
• 11 one-on-one online interviews in Cantonese conducted by a research assistant unconnected to teaching, transcribed verbatim, and analyzed by thematic analysis (Braun & Clarke, 2022); three themes-plus concerned task design, valuing the learning opportunity, self-confidence, plagiarism fear, and Teaching approaches.
• Instructional and declaration artifacts: course outlines, in-class briefing slides, and all 158 students' self-declarations of GenAI use submitted with their assignments, analyzed by qualitative content analysis.
• Context of policy: the institution allows four levels of GenAI adoption (no use; limited use with permission; use with acknowledgment; free use); all three courses operated at level two, but individual instructors set and communicated their own rules, and Hong Kong universities had briefly banned GenAI entirely in early 2023.
• Analysis: descriptive statistics, Mann–Whitney U tests (non-parametric, chosen over t tests because data were not normally distributed), and iterative coding; findings rest on self-reported data, a single institution, and the low Course C response rate.

What this means for practice

• Instructors. Do not assume a permissive GenAI policy communicates itself: nine of the 11 interviewees described the allowance as a potential "trap", and only 37.6 percent of surveyed students (32 of 85) used GenAI in their assessments at all.
• Designers. Show worked examples of acceptable and unacceptable use and give timely feedback on declarations instead of restating rules, since declarations diverged from survey answers in every course (Course A 8 declared vs 14 surveyed; Course B 6 vs 10; Course C 12 vs 8).
• Instructors. Anchor assessment tasks in reflective, experience-based professional problems: non-adopters most often preferred to work alone (41 of 53, 77.4 percent) or feared being accused of plagiarism (22, 41.5 percent), while only 7 (13.2 percent) said they lacked the knowledge or skills to use GenAI; build confidence and task value deliberately, since Scaffolding and feedback that strengthen competence and value are assessment-design levers, not adjuncts.
• Administrators. Harmonize GenAI rules across a program, because student behavior tracked the wider institutional culture — including a one-month institutional ban in 2023 and inconsistent rules in other courses — rather than any single course brief; use teacher education assessments to grow future teachers' assessment literacy, since embedding GenAI in them gives first-hand experience of responsible use that participants otherwise showed little intention of carrying into their own school assessments.
• Instructors. Measure actual uptake and treat non-use as a legitimate choice: assessment engagement averaged 4.21 out of 5 with no significant difference between adopters and non-adopters (U = 781.5, p = 0.536), so tool use is not what carried engagement.

Limitations

• The study covers three purposively sampled courses at a single research-intensive Hong Kong university, with a survey of 85 of 158 enrolled students and response rates of 100 percent in Course A, 60 percent in Course B, and 32.9 percent in Course C.
• Findings rest on students' self-reported data, and reported use did not match behavior: declarations submitted with assignments disagreed with anonymous survey responses in every course (Course A 8 declared vs 14 surveyed; Course B 6 vs 10; Course C 12 vs 8).
• All three courses operated at the same permitted level of the institution's four-level GenAI policy, with individual instructors setting their own rules, so the "trap" perception is bound to a context that had briefly banned GenAI entirely in 2023.
• The qualitative themes come from 11 Cantonese-language interviews at that one institution and cannot be generalized to policy settings elsewhere.

Citation

Zou, T. X. P., Hounsell, D., Huijser, H., & Tse, R. T. L. (2026). "Is this a trap?": Student teachers' perceptions and adoption of GenAI in assessments in three teacher education courses. Australasian Journal of Educational Technology, 42(1), 1–17.