Role of generative AI literary assistants in enhancing ninth-grade students' writing motivation, flow and achievement

Synthesis: Wang, Wang, and Liu (2026) use a quasi-experimental, single-group pre–post mixed-methods design to examine the multifaceted effects of integrating Generative AI (GAI) literary assistants into ninth-grade Writing instruction, focusing on students' writing motivation, flow experiences, and overall writing performance. Across a three-day GAI writing camp, students used ChatGPT to emulate the distinctive styles of four literary figures (Su Shi, Yu Kwang-Chung, Xi Murong, and Jian Zhen) through structured Prompt Engineering, a deliberate shift from generic AI use to literature-informed Prompt Engineering. While general writing motivation showed no statistically significant increase (pre-test mean 2.93 → post-test 3.00, t = −0.59, p = .560), flow improved significantly (3.15 → 3.60, t = −2.86, p = .007, Cohen's d = −0.45) and writing performance rose markedly (77.13 → 81.40, t = −5.26, p < .001, d = −0.83). Qualitative data revealed a marked surge in task-specific engagement and situational motivation, particularly during the revision phase. The authors also critically examine the risks of Over-Reliance on AI, which could diminish engagement in areas requiring creative expression and Critical Thinking.

Key Findings

Design and sample. A quasi-experimental, single-group pre–post design with mixed-methods data collection was used with ninth-grade students from a junior high in northern Taiwan. From 44 recruited students, the final valid sample was 40 after excluding participants for failing to complete at least two-thirds of camp activities (N = 1), a questionnaire missing-response rate over 20% (N = 2), and a pre–post writing-score change exceeding two standard deviations (N = 1).

Literature-informed prompt engineering. The core pedagogical innovation moved students from undirected AI use to structured prompts that emulated four literary voices — Su Shi (bold, philosophical), Yu Kwang-Chung (emotional, lyrical), Xi Murong (delicate, youthful), and Jian Zhen (female perspective) — aligned with the attention and relevance components of Keller's ARCS-V Motivation model. Interactions ran in traditional Chinese using GPT-4 (temperature 0.7, top-p 1.0, max 200 tokens) for style-based revision, emotional/rhetorical enhancement, structural review, and summary generation.

Improved writing quality (large effect). Writing achievement rose from a pre-test mean of 77.13 to 81.40 (t = −5.26, p < .001, Cohen's d = −0.83, 95% CI [−1.19, −0.47]; strong pre–post correlation r = 0.75, p < .001), with gains in content completeness, language expression, and organizational structure. Improvement in Creativity and originality was comparatively limited, indicating AI most strongly boosts technical writing skills.

Flow experience improved (medium effect). Average flow rose from 3.15 to 3.60 (t = −2.86, p = .007, Cohen's d = −0.45). Students readily entered flow during the revision phase, describing an "immersive interactive experience"; however, some reported decreased flow when excessive AI support reduced the perceived challenge and sense of full immersion.

Situational vs. general motivation. General writing motivation did not change significantly (2.93 → 3.00, t = −0.59, p = .560; pre–post correlation 0.00), while qualitative data showed task-specific engagement and situational motivation surged, especially during revision — pointing to the context-dependence of AI's motivational effects and aligning with Self-Determination Theory Theory.

GAI as cognitive scaffold rather than motivational trait. Post-intervention comparisons showed writing performance exceeding flow, and flow exceeding motivation, across the group (e.g., all 40 students' writing-performance scores exceeded their flow scores, Z = −6.17, p < .001; 31 of 40 had motivation below flow, Z = −3.32, p < .001). This suggests GAI primarily scaffolds writing outcomes and task engagement rather than reliably raising stable intrinsic motivation.

Developing authorship and Self-Regulated Learning. Students selectively accepted AI suggestions, distinguishing "my tone" from AI output (e.g., "ChatGPT made it too ornate, I wouldn't write like that myself"), signaling emerging authorial voice. Post-test Writing Self-Efficacy scores increased in ideation ability, rhetorical application, and self-AI Regulation in Education, and a 20-point, four-dimension scoring rubric (content, organization, language, creativity, 25% weight each) anchored the writing Assessment.

Over-Reliance risk and pedagogical recommendation. The authors caution that over-reliance on AI could diminish engagement in areas requiring creative expression and critical thinking, concluding with scaffolded pedagogical recommendations for integrating GAI writing assistants while preserving students' creative and critical capacities — directly relevant to the knowledge base's Writing and Over-Reliance concerns.

What this means for practice

• Instructors. Replace generic AI prompts with structured, literature-informed Prompt Engineering—students writing through the voice of a named literary figure—which is the design the authors credit for the flow and performance gains in this camp.
• Educators. Time AI support to the revision phase, where students most readily entered flow, and pull it back when it removes the perceived challenge that sustains immersion.
• Instructors. Ask students to state what is theirs and what is the AI's ("ChatGPT made it too ornate, I wouldn't write like that myself") to build authorial voice and self-regulation while using the tool.
• Educators. Score with a four-dimension rubric (content, organization, language, Creativity) and expect the weakest return on originality, since improvements concentrated in technical writing skills.
• Educators. Do not rely on GAI to raise stable Motivation: general writing motivation did not move in this study while situational engagement surged, so design for task-level engagement rather than assuming a lasting motivational effect.

Limitations

• A quasi-experimental single-group pre–post design with no control group, so the writing performance rise (77.13 → 81.40) cannot be separated from practice effects, novelty, or the camp format itself.
• Forty students from one junior high in northern Taiwan, after four of the 44 recruits were excluded—a single-site, three-day intervention that is far shorter than a normal instructional cycle.
• Motivation, flow, and self-efficacy were measured by self-report questionnaires; general motivation showed no significant change (2.93 → 3.00, t = −0.59, p = .560) with a pre–post correlation of 0.00, so those scales may not have been sensitive in this design.
• The design cannot isolate the literary-character simulation from GAI use in general, and does not test whether the gains hold once the novelty of the camp fades.

Citation

Wang, S. B., Wang, S. I.-C., & Liu, E. Z.-F. (2026). Role of generative AI literary assistants in enhancing ninth-grade students' writing motivation, flow and achievement. Computers and Education Open, 100339. https://doi.org/10.1016/j.caeo.2026.100339