A longitudinal constructionist case study of five student positions on AI as a creative collaborator in open source AI art pedagogy

Technical Summary: A Longitudinal Constructionist Case Study of Five Student Positions on AI as a Creative Collaborator

Problem Statement
Constructionist learning theory posits that pedagogical environments designed with "low floor, wide walls, and high ceiling" produce individually differentiated student trajectories, where this differentiation is itself the pedagogical achievement. While this is well-established regarding the diversity of student artifacts (projects, code, creative directions), it remains under-tested at the granularity of a structured spectrum of intellectual stances on a focal technology. Furthermore, existing literature rarely tracks these stances longitudinally into post-graduation practice. This paper addresses the gap in understanding whether a single constructionist seminar on generative AI can produce a defensible spectrum of student positions regarding AI's role in creative authority, and whether these positions persist after graduation.

Methodology
The study employs a qualitative multiple case study design (Yin, 2019) with embedded units of analysis, focusing on a single cohort of five students from a two-year zemi (Japanese seminar structure) at Kyoto Seika University's Faculty of Media Creation. The observation window spanned approximately 25 months, from the seminar's inception in April 2024 through the March 2026 graduation exhibition and into early post-graduation practice (April–May 2026).

• Pedagogical Context: The pedagogy was explicitly constructionist by design, anchored in the first semester by ComfyUI (an open-source, node-based interface for generative AI) to ensure pipeline transparency. The tool stack subsequently expanded to include AI-assisted code editing (Cursor), proprietary video generation, LLMs, and game engines based on individual project needs.
• Data Sources: The analysis triangulates five streams of qualitative evidence:
  1. Longitudinal Interviews: Asynchronous email interviews (Round 1 and Round 2) conducted in April–May 2026 with five students (pseudonyms: XJ, LX, LY, LJ, WS).
  2. Baseline Survey: An end-of-instruction qualitative survey (October 2024) capturing initial metaphors and help-seeking dispositions.
  3. Production Notes: Student-authored academic documents (9–12 pages) submitted for graduation oral examinations.
  4. Supervision Records: Approximately 1,275 instructor-student exchanges via Discord, including long-form direction letters.
  5. Practitioner Documentation: Reflective logs and teaching documents authored by the researcher (who was also the seminar instructor).
• Analytic Approach: Reflexive thematic analysis (Braun & Clarke, 2006) was conducted by a single researcher. The analysis prioritized within-case familiarization before cross-case synthesis, treating researcher subjectivity as an analytical resource.

Key Contributions
The paper makes five substantive contributions to the literature on generative AI in higher education and arts pedagogy:

1. Five-Position Spectrum: It documents a structured spectrum of five distinct stances on "AI as a creative collaborator," ranging from "expansive system-designer" to "bounded-philosophical," providing granularity previously absent in AI-art pedagogy literature.
2. Within-Cohort Triangulation: It presents an independently constructed mental-model evolution typology that converges one-to-one with the five-position spectrum, validating the findings across two different analytic paths (trajectory vs. mental model) and data sources.
3. Asymmetric-Amplification Mechanisms: It identifies four candidate mechanisms by which AI scaffolding succeeds at the conceptual layer but fails at the realization layer, each anchored to a specific student case and corroborated by their baseline help-seeking dispositions.
4. Multi-Modal Register-Mix Mentorship: It proposes a typology of seven uptake modes, demonstrating that students often operate multiple mentorship modes simultaneously across different content layers (conceptual, technical, implementation) and supervision contexts.
5. Methodological Argument: It argues that pairing an end-of-instruction baseline survey with longitudinal follow-up interviews, triangulated against practitioner records, yields a "cross-validation dividend" that interview-only or record-only analyses cannot achieve.

Results
The study identifies five distinct student positions on AI as a creative collaborator:

1. Expansive System-Designer (XJ): Views AI as a co-designer of a tightly coupled human-AI generative pipeline. The student's mental model evolved via cumulative-pluralisation, enriching a single "cooking recipe" metaphor into multiple layered metaphors.
2. AI-as-Operations-Manual (WS): Views AI as knowing how the software operates, but maintains strict human creative authority. The mental model remained a sustained metaphor (culinary register) with stable conceptual framing.
3. Conceptual-Amplifier-Only (LJ): Relies on AI for conceptual and linguistic work but excludes it from technical execution. This case exhibited an acquired metaphor, where substantive mental-model articulation emerged only after the 2024 baseline, facilitated by a multi-supervisor handoff.
4. Bounded-Ethical-and-Pragmatic (LX): Limits AI use due to ethical authorship hesitation (viewing AI as a creative agent deserving credit) and engineering pragmatism (inefficiency of AI-generated code). The mental model shifted from a tentative metaphor to a literal-pipeline articulation, rejecting metaphor for engineering precision.
5. Bounded-Philosophical (LY): Limits AI use based on the belief that creative authority requires embodied human experience. The mental model shifted from a social-structural metaphor ("boss and employees") to a technical-precision articulation regarding latent space convergence.

Asymmetric Amplification and Workarounds
The study found that AI scaffolding often amplifies conceptual understanding but fails at realization through four mechanisms:

• Modality/Training Limits: AI fails when tools are GUI-heavy with sparse training data (LJ).
• Directional Narrowing: Students with high capability may reject practitioner advice for more ambitious realization (LX).
• Medium Affordance Ceilings: Students may choose mediums that compromise aesthetic impact to serve a specific creative vision (LY).
• Disclosure Limits: Students who do not surface technical details prevent iterative advice, leading to realization gaps (WS).

Three workaround paths were identified: AI co-engineering with practitioner scaffolding, practitioner-directed high-level architecture with research handoff, and student self-directed template modification.

Significance and Claims
The paper claims that the constructionist "wide walls" environment does not merely produce diverse projects but fosters a structured spectrum of intellectual stances on the medium. However, the authors modestly frame this outcome not as a direct product of the pedagogy alone, but as the result of the pedagogy providing a space where multi-modal register-mix mentorship operates to amplify and articulate existing student capacities.

The study challenges the "homogenization" hypothesis (that AI leads to uniform outputs) by demonstrating that under sustained constructionist mentorship, the same focal technology produced systematically differentiated trajectories. It further argues that the differentiation is not merely an amplification of pre-existing individual differences but is constructed through the seminar as a formation site, evidenced by students explicitly attributing their evolving stances to the pedagogical experience.

The paper concludes that these findings offer a calibrated reading of constructionist theory for the AI era and provide a transferable methodological resource for practitioner-inquiry case studies in higher education, specifically the value of triangulating longitudinal interview data with baseline surveys and practitioner records. The authors note that generalization beyond this single cohort requires further testing with a parallel postgraduate cohort at the University of Macau.