Instructor-Designed AI Tutors in University Foreign Language Education: A Mixed-Methods Study of Learner Motivation and Reflective Learning Experience Based on Self-Determination Theory

Synthesis: Lee and Kwon (2026) report a mixed-methods study of an instructor-designed, customized Japanese-language GPT tutor in a South Korean university general-education course, drawing on self-determination theory (SDT) and the noticing hypothesis. With 74 undergraduates, they find that continuous use of the instructor-built AI tutor was associated with high satisfaction of autonomy and relatedness needs, three distinct cognitive-noticing experiences, and learner perceptions of the tool as a structured learning environment rather than a convenience tool. The central claim is that the educational effectiveness of generative AI in foreign language learning depends less on frequency of use than on the quality of pedagogical design underlying its deployment.

Key Findings

• Autonomy-supportive design. Learners reported high perceived helpfulness for self-directed learning (M = 4.39, SD = 0.70) and a positive change in learning motivation (M = 3.95). Autonomy was realized within an instructor-designed system aligned to course objectives, not through general-purpose ChatGPT use alone.
• Production-oriented competence gains. Competence satisfaction was comparatively lower overall (early-stage learners did not yet perceive performance gains), but domains tied to production — writing, vocabulary, sentence generation — showed relatively higher means, and autonomy perceptions correlated most strongly with production-oriented activities.
• Relatedness through affective safety. Immediate feedback (78.4%) and affective safety (60.8%) were the most-endorsed relatedness indicators. The instructor-designed GPTs functioned as nonjudgmental, immediately responsive practice spaces for learners anxious about making errors in class.
• Three noticing experiences (RQ2). Qualitative analysis identified (1) AI-supported clarification of linguistic form, (2) noticing through intentional error generation plus AI feedback, and (3) metacognitive regulation of learning strategies.
• Design over frequency. Out-of-class usage frequency showed no significant correlation with SDT variables — how the tool is used and under what design conditions matters more than how often.
• Reconfigured learning mode (RQ3). Learners shifted from passive memorization/answer-confirmation toward active, interaction-based, output-oriented learning, and attributed the psychological safety they felt to the instructor's pedagogical and relational design (e.g., one learner contrasted "a rational type" general GPT with the instructor's "very bright" customized GPT).

What this means for practice

• Instructors. Build course-scoped AI tutors tied to the course objectives, learner proficiency, and a curated knowledge base rather than pointing students at general-purpose chatbots.
• Instructors. Operationalize the noticing hypothesis with a three-stage sequence: activate prior knowledge, have learners deliberately generate errors, then produce context-based sentences and work through the feedback that follows.
• Instructors. Design for affective safety — nonjudgmental, immediately responsive interaction — because the relatedness gains came from instructional design, not from the technology itself.
• Instructors. Deploy the tutor in a complementary role alongside existing methods, since learners cited usage limits and preferences such as handwriting-based study as real constraints.
• Faculty developers. Support instructors in designing AI around autonomy, competence, and relatedness, because effectiveness tracked design quality rather than how often learners used the tool.

Limitations

• Single-course scope: 74 consenting students in one general Japanese course at one South Korean university, followed across one semester with no control or comparison condition.
• All motivation and noticing data came from an end-of-semester self-report survey; autonomy used two single-item indicators and relatedness a multiple-response item, and optional open-ended items had varying numbers of valid responses (n = 54–68 across competence sub-domains).
• Competence was self-assessed across six sub-domains (means 2.67–2.94), not measured by an objective proficiency test.
• The design is correlational — out-of-class usage frequency showed no significant correlation with the SDT variables — so no causal claim about motivation or learning gains is supported.

Citation

Lee, H., & Kwon, H. (2026). Instructor-Designed AI Tutors in University Foreign Language Education: A Mixed-Methods Study of Learner Motivation and Reflective Learning Experience Based on Self-Determination Theory. Trends in Higher Education, 5(3), 78.