AIEd Knowledge BaseAIEd KB
On this page

Research Article

The Influence of Generative AI Usage Styles on Creative Self-Beliefs: Findings from a Longitudinal Study in Design Education

added 2026-09-18 · updated 2026-09-19

Synthesis: Okamoto and Inasaka ask whether generative AI use changes Creative Self-Beliefs (CSB) in design education, and whether any effect depends on how AI is used rather than whether it is used at all. Sixty-four third-year students in a Display Design Theory and Practice course completed three hand-drawn design sprints under three conditions, no AI (Assignment 1), mandatory AI (Assignment 2) and optional AI (Assignment 3), with CSB measured at four time points (T0–T3) as Creative Self-Efficacy (CSE) and Creative Personal Identity (CPI). Using AI was not associated with CSB gains: mean CSE moved from 3.79 to 3.94 and CPI from 4.60 to 4.77, but a mixed-effects model found no significant time, group or interaction effects on CSE and only a marginal T3 effect on CPI (β = 0.34, p = .063). Style mattered instead: students who valued asking AI for feedback (ΔCSE +0.72, p = .018) or summarizing meaning (+0.88, p = .031) gained more, while delegating idea generation to AI did not separate them at all. Students overwhelmingly read AI as a way to finish tasks faster, not as a creativity aid. The takeaway is that AI's route to creative self-belief runs through reflective, feedback-oriented use rather than through handing creative work over.

Key Findings

  1. The sample is small, one course, four waves. One hundred and twenty third-year undergraduates took the course, but only the 64 who completed all four questionnaires (T0 before Assignment 1, then T1–T3 after each sprint) entered the analysis, split at T3 into an optional-use group (n = 30) and a non-use group (n = 34).
  2. AI use on its own predicted nothing. In the mixed-effects model with participant random effects, CSE showed no significant main effect of group (β = −0.25, p = 0.48), of time (T1–T3 all p ≥ 0.26) or of their interaction; CPI showed only a marginal T3 main effect (β = 0.34, p = .063) with a non-significant group effect (β = 0.03, p = 0.94).
  3. Gains were explained by baseline, not by AI. Regressing change scores (T3−T0) on AI use and baseline showed AI use non-significant (ΔCSE β = −0.22, p = .40; ΔCPI β = −0.35, p = .15), while baseline T0 scores carried significant negative coefficients (β = −0.40 and −0.56, both p < .001), a ceiling effect rather than a treatment effect.
  4. Descriptively, belief rose after the phase when AI was optional. Mean CSE was 3.79 (T0), 3.82 (T1), 3.72 (T2) and 3.94 (T3), and mean CPI 4.60, 4.57, 4.57 and 4.77, with the dip at T2, when AI use was compulsory, instead of the rise.
  5. Non-users gained more than users. Non-users rose from CSE 3.90 to 4.11 and CPI 4.59 to 4.93 across T0–T3, while AI users moved only from 3.66 to 3.74 on CSE and stayed flat on CPI (4.61 to 4.59); users also started lower, suggesting students with weaker initial CSE were likelier to take up AI when it became optional.
  6. Students valued AI for ease, not for creativity. A Mann–Whitney U comparison found a significant difference only for "I find it easier to proceed with tasks when using ChatGPT" (4.12 vs 5.17, W = 284.5, p = .0017); perceived usefulness for one's own creativity (4.50 vs 4.90, p = .249) and intention to continue use (p = .320) did not differ.
  7. Idea generation dominated the reported usage styles. At T2, participants picked idea generation 69.9%, organizing premises 36.9%, research 31.1%, question generation 25.2%, requesting feedback 23.3%, summarizing meaning 19.4% and role-playing 9.7%; among T3 users idea generation stayed top at 66.7%, role-playing rose from 11.1% to 23.1% and question generation fell from 25.0% to 5.1%.
  8. Feedback-seeking predicted persistence with AI. Comparing T2 usage categories against continued use at T3 with Pearson's chi-square, only "requesting feedback" was significant (χ²(1) = 4.35, p = .037), selected by 36.1% of continued users against 14.8% of those who stopped.
  9. Two usage styles separated the CSE gains. Among T3 users, ΔCSE was higher for those who rated requesting feedback useful (+0.72, Cliff's δ = 0.50, p = .018, n = 11) and summarizing meaning (+0.88, δ = 0.57, p = .031, n = 6), with every other style non-significant; for ΔCPI only idea generation (+0.71, p = .052) and summarizing meaning (+0.77, p = .080) reached marginal levels.

How the study distinguishes AI usage styles

The design rests on the argument that "use versus non-use" is the wrong independent variable for a technology that can be an idea generator, a search substitute, a feedback partner or a meaning-making interlocutor. The authors measure styles with a multiple-response item asking which of seven uses students found helpful, labeled A to G: idea generation, organizing premises, role-playing, question generation, requesting feedback, research and summarizing meaning. Each option maps to an AI usage guideline distributed in class before Assignment 2, so the categories are the Pedagogies and Teaching Strategies's own vocabulary rather than a survey scale.

CSB itself is measured with an established eight-item instrument, three items for CSE ("I am good at coming up with new ideas") and five for CPI ("I think I am a creative person"), each on a 7-point Likert scale. A separate three-item attitudes scale taps perceived usefulness for creativity, ease of task progression and intention to continue. Because the authors work from a mini-c view of creativity as personally meaningful interpretation rather than socially recognized production, they read a framing statement about everyday creativity to students before every wave, following evidence that framing shifts whether creativity is seen as malleable.

The longitudinal design and sample

The setting is a single third-year course at Chiba Institute of Technology in which students ran design sprints: thirty minutes generating textual idea seeds, sixty minutes developing at least five into concepts with visual representations, then refining one into a presentation panel, all hand-drawn. Assignment 1 (October 2023) prohibited AI, Assignment 2 (23 October 2023) required it, and Assignment 3 (11 December 2023) left it optional. This sequencing is what lets the paper separate mandatory exposure from voluntary uptake, and it is why belief dips at T2 and rises at T3 in the descriptive series.

Attrition is substantial and structural: 120 students participated, 64 completed all four surveys, so the analyzed picture is the committed half of the cohort. Analyses run in R include mixed-effects models with participant random effects, the Mann–Whitney U test for attitudes, McNemar's test for within-user change in usage styles between T2 and T3, chi-square for style-to-continuation links, and per-style comparisons of change scores with Cliff's δ as effect size. Qualitative free-text responses at T3 supplement the numbers.

Results: AI use versus AI usage style

The study's headline is a null on use and a signal on style. Whether a student used generative AI, and whether the class was required to, did not explain movement in creative self-efficacy or creative personal identity; the only near-miss was a marginal CPI main effect at T3. Within the subgroup of users, however, the students who treated AI as something to interrogate rather than something to consult gained more CSE. This is the paper's Human AI Collaboration finding in concrete form: AI positioned as an evaluator of one's own ideas sits differently in the creative process than AI positioned as a generator of them.

What separates the styles is the role the student keeps rather than how much work is handed over: the association is with interrogating the output, not with delegation as such. Idea generation, the most-used style by a wide margin, produced no CSE difference (+0.11, p = .41) and only a marginal CPI difference, and delegating the generative part of the work to AI is precisely what the students who stopped using it described resisting, with comments such as "I wanted to try doing it on my own" and "I wanted to test my own creativity". The Feedback and summarizing styles, by contrast, keep the student as the author of the idea and put AI in the role of revisiting, reinterpreting and strengthening it.

Interpretation: ease of use is not creative self-belief

The authors read the attitude results as the interpretive key. Students endorsed AI as making tasks easier far more strongly than they endorsed it as improving their own creativity, so a productivity gain and a self-belief gain appear to be different constructs that the same tool affects unequally. That reading connects to Bandura's account of Self-Efficacy: an outcome supports efficacy beliefs only when it is attributed to one's own action and felt to be under one's control, and a tool that behaves indeterminately and produces output automatically makes that attribution harder. Drawing on Ueno's account of tool use, they suggest generative AI had not yet been internalized as a tool in the way a hammer is, so its output was not internalized as the user's own creativity.

The causal caution is explicit. The style–CSE associations rest on 6 to 11 students per category, the study is one course and roughly eleven weeks, and the design cannot separate the effect of a style from the disposition of students who choose it. Their preferred framing is that AI functions as a supportive element under specific relational conditions rather than a cause of change, and that CPI is simply less malleable in the short term, consistent with earlier longitudinal work showing CSE and CPI change slowly and reciprocally.

What this means for practice

  • Educators. Treat what students do with AI, not whether they use it, as the lever: banning, requiring and permitting AI were all weaker than shaping what students do with it.
  • Educators. Teach the styles tied to gains deliberately — have students put their own draft in front of AI and ask what it undermines, and have them restate and re-interpret their material — instead of treating AI as a faster search box or an idea vending machine.
  • Learning designers. Build Learner Identity work — helping students see themselves as the author of the idea — into AI integration alongside tool training.
  • Educators. Read students who opt out for reasons of creative ownership as a signal worth listening to rather than a compliance problem.

Limitations

  • Attrition is structural: 120 third-year students took the course but only the 64 who completed all four questionnaires entered the analysis, so the picture is the committed half of one cohort, split at T3 into an optional-use group (n = 30) and a non-use group (n = 34).
  • Every measure is self-report, collected as eight 7-point Likert items (three for CSE, five for CPI) plus a three-item attitudes scale, administered four times across a single design course at Chiba Institute of Technology between October and December 2023, with real but limited transfer to other disciplines and levels.
  • The style–CSE associations rest on very small subgroups — 6 to 11 students per usage category (requesting feedback n = 11, summarizing meaning n = 6) — and the per-style analyses are exploratory rather than confirmatory.
  • The design is correlational and the window of about eleven weeks is too short for the human–AI relationship to have stabilized, so it cannot separate the effect of a usage style from the disposition of the students who chose it; the authors themselves frame AI as a supportive element under specific relational conditions rather than as a cause of change, and their forward agenda is longer longitudinal work that adds perceived control and sense of agency over AI, with the wider caution — captured in work on AIED limitations — that output quality and self-perception are not interchangeable outcomes.

Connected Concepts

  • Creativity — the outcome domain, framed through the mini-c view of personally meaningful novelty
  • Self-Efficacy — CSE and creative personal identity as the two measured components of creative self-belief
  • Student Engagement — the persistence question the paper asks about sustained voluntary AI use
  • Design Thinking — the design sprint pedagogy and hand-drawn workflow in which the study ran
  • Design Education — the single-course disciplinary setting and its transferability limits
  • Generative AI — the technology whose use styles were compared
  • Large Language Models (LLMs) — conversational systems such as ChatGPT supplying the usage options
  • Human AI Collaboration — AI as feedback partner and meaning-making interlocutor rather than search substitute
  • Learner Identity — creative personal identity and the attribution of output to oneself
  • Learner Agency — perceived control over the creative process as the proposed explanatory mechanism
  • Feedback — requesting feedback from AI as the usage style most clearly tied to CSE gains
  • Limitations in AIEd Research — small subgroups, attrition and correlational design in AIED evidence

Connected Articles

Citation

Okamoto, R., & Inasaka, A. (2026). The Influence of Generative AI Usage Styles on Creative Self-Beliefs: Findings from a Longitudinal Study in Design Education. Manuscript, Saitama Institute of Technology and Chiba Institute of Technology.

Metadata

Study designLongitudinal Study, Survey
DisciplineDesign Education
LevelHigher Ed
Intended audienceResearchers, Instructors, Instructional Designers
Learning and instructionCreativity, Self-Efficacy, Student Engagement
Technologies and techniquesGenerative AI, Large Language Models (LLMs)
Research methods and evaluationQuantitative Research
Foundations of AI in educationDesign Thinking, Human AI Collaboration, Learner Identity
Page kindEvaluation
SourceHistoryEditChat

Embed this page

Copy the code below to embed a chromeless version of this page in a learning management system or other website. The embedded view hides the site header, navigation, and footer.