Hong Kong TESOL trainees used generative AI selectively across 10 multimodal website projects

Moorhouse, Hafner and Ho examine how prospective language teachers used generative AI while creating digital multimodal projects in a Hong Kong MA TESL course. The 13-week course asked groups to build a website as a capstone task, with AI use optional rather than required. Students received a three-hour workshop, submitted projects in week 12 and commented on peers' work in week 13. The study asks not merely which tools appeared, but how participants decided when AI supported or weakened their purposes as multimodal composers and future teachers.

The course enrolled 137 students; 22 consented to the research and represented 10 group projects. Nineteen participants were women and three were men, all aged 22 to 28 and from mainland China. Most reported little or no teaching experience and expected to teach in mainland China, Hong Kong or Macao. Evidence comprised 10 individual or group stimulated-recall interviews lasting about 50 to 65 minutes and the 10 completed websites. Interviews were conducted in Mandarin by the third author, who was not the course teacher, and machine-assisted transcription or translation received human oversight.

All 10 groups used generative AI selectively. Nine used large language models, two used image generators and two used video generators; every group also used multimodal or website-building tools. Across creation activities, eight groups generated images, seven summarized material, six produced data visualizations, six analyzed data and four generated text. Eight groups sought ideas or advice from AI and seven sought feedback. These counts show a broad repertoire rather than uniform dependence: groups combined tools differently and sometimes rejected an AI contribution when it conflicted with the intended message or their desired level of authorship.

Participants described decisions involving efficiency, perceived capability, authenticity, meaning-making and degree of personal involvement. They reported benefits such as confidence, creativity, digital literacy and critical reflection, alongside concerns about reliability and contextual fit. Those are valuable accounts of experience, not objective gains. The study has no comparison group, pre-post assessment, blind scoring or measure of later classroom transfer. Only 22 of 137 students volunteered, all came from mainland China, and the single postgraduate course cannot represent all Hong Kong trainees, practicing teachers or language-learning contexts.

For Hong Kong teacher education, the study offers a concrete curriculum pattern. Programmes can teach multimodal design and AI evaluation together, require students to record which suggestions they accepted or rejected, and assess whether words, images, charts and navigation serve a coherent teaching purpose. A reflection can distinguish assistance with routine production from decisions that require disciplinary, cultural and pedagogical judgment. Bilingual reviewers can check English, Chinese and local classroom context, while protected student or placement data should remain outside unapproved tools.

The defensible finding is that these participants exercised selective agency across 10 authentic projects; the paper does not prove that generative AI improved teacher competence or learner outcomes. Future work could compare scaffolded and unscaffolded cohorts, score products blind to condition, examine delayed independent composing and follow graduates into classroom practice. Meanwhile, the study gives Hong Kong educators a useful assessment target: not the number of AI features used, but the quality of the human reasoning that determines when a generated contribution belongs in a purposeful educational design.