On this page

Synthesis: A 16-week graduate Training and Development Seminar in Taiwan measured self-directed learning and course motivation before and after a semester of AI-supported professional training work. Both totals rose across all responses (self-directed learning 80.62 to 87.38; ARCS motivation 98.08 to 106.54), and eight matchable records gained 7.25 and 9.88 points on average (exact p-values .015625 and .03125). Reflections describe AI as an always-available structuring aid with obvious professional relevance, alongside worry about dependence and self-doubt. With no comparison group, it documents plausible change in one seminar, not an effect of AI support.

Key Findings

  1. Both self-directed learning and motivation rose, but the design cannot say why. Course content, peers, instructor attention, novelty, and repeated testing all remain plausible explanations.
  2. The inference rests on eight matched pairs. Gains were consistent across the group, but a sample that size cannot separate the AI support from everything else that happened in the seminar.
  3. Near-uniform movement inflates the effect sizes. Almost every participant moved in the same direction, which makes the rank-based effect look maximal even though the intervals rest on eight people.
  4. The motivation total measures reaction to the course, not independent motivation. Its items ask about teaching, content, feedback, and grades, and the self-directed learning questionnaire carries motivational items of its own.
  5. Students valued generative AI as an always-available structuring aid, especially for organizing material and monitoring progress between sessions.
  6. Human review remained the gate. Students described accepting AI output only after a knowledgeable reader checked it, which is delegated rather than surrendered agency.

What the numbers can and cannot support

An elective seminar meeting three hours per week over 16 weeks in academic year 2024–2025 drew 13 graduate students, 10 master's and three doctoral. Measurement used a 20-item SDLI-based questionnaire (20–100) and a 24-item ARCS questionnaire (24–120), with alpha between .864 and .941. Only eight email addresses matched pretest records of 11 September 2024 to posttest records of 12–17 December 2024; five records at each occasion could not be linked, so inference used eight, description all 13.

The proposal planned action research, but no dated cycle record survives; the article reports itself as exploratory single-group pretest–posttest. Its outcome is self-directed learning, not self-regulation: an SDLI total says nothing about whether SRL components changed. Fan et al. (2025) found ChatGPT-assisted writing improved essays without matching transfer or intrinsic motivation gains; Bastani et al. (2025) showed practice-time assistance diverging from later unaided performance; Deng et al. (2025) synthesized 69 experimental articles with no significant overall effect on self-efficacy. No unaided transfer test was run.

What students said about support and doubt

The 14 extracts from nine participant codes, dated Weeks 7, 9 and 16, with eight from Week 16, describe availability and professional reach: one student listed Perplexity, Claude, ChatGPT, Gamma, Suno and Napkin as tools tried on reports and data analysis, one student's report rather than an assigned list. The author will not claim anyone enacted a full regulation cycle; the Week 7 excerpts already show these ideas.

The qualifying theme is sharper. One participant described self-doubt on noticing AI output could read as more creative than her own work. Another recounted a group task where students were told "not to type a single word ourselves" and accepted AI-only output because a knowledgeable gatekeeper reviewed it. Satisfactory work under time pressure is not evidence of understanding.

Assessing what remains after the tool is removed

The paper's most usable contribution is terminological. AI support means engaging with generated content; it is not generative learning in Wittrock's or Fiorella and Mayer's sense, where the learner builds the connections, so a course that only inspects deliverables will not notice the difference.

Proposals for revised teaching, not a record of this seminar: students state a goal and initial judgment before using a tool, compare output with credible sources or a rubric, explain what they accepted or rejected, and say what they could do unaided. Assessment would pair assisted products with an oral explanation or a critique of an erroneous output, making scaffolding visible rather than assumed. Graded reflection should be free to criticize the AI and the course.

What this means for practice

  • Instructors: Do not read end-of-course totals as evidence that AI helped; require a goal and initial attempt before tool use so planning leaves a trace.
  • Instructors: Keep one assessment moment unaided, because assisted performance and independent understanding can diverge.
  • Designers: Build the gatekeeper role into task instructions; the student who accepted AI-only group work did so because a knowledgeable reviewer checked it.
  • Researchers: Use stable identifiers. Losing five of 13 records to email matching left a paired sample of eight, all women.

Limitations

  • Single group, no randomization, comparison or follow-up, so durable learning goes untested.
  • Selective matching: eight of 13 records, all women and seven master's students, so the paired subset may differ from unmatched respondents.
  • Self-report with alpha as the only reliability evidence; the SDL grouping merges planning with monitoring, and duplicate item columns needed a scoring rule.
  • Instructor and researcher were the same person with no grading safeguard, and formal ethics approval was not obtained on the authors' own determination.

Citation

Chen, Y.-C. (2026). Generative AI-supported learning, self-directed learning, and motivation in a graduate seminar: an exploratory single-group study. Frontiers in Education, 11, 2008734.

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.