Leveraging ChatGPT in academic writing: ChatGPT enhances students' writing quality, writing experience, and ownership

Synthesis: Bauer, Fütterer, Brucker, and Gerjets (2026) test whether ChatGPT promotes students' Writing skills — especially when writing without ChatGPT after having used it. In an ABA-within-subjects experiment lasting three hours, 21 first-year international fashion-design students in Berlin (admitted via IELTS/TOEFL performance, mean English proficiency 4.67/6) wrote introductions to three scientific papers on permuted fashion-related topics: first without AI, second with ChatGPT (via a GPT-3.5 API embedded in the experiment interface), and third again without assistance, each in 35–50 minutes. ChatGPT produced statistically significant gains in writing quality both during use (examiner total text scores d = 1.18, p < .001; readability d = 1.09; cohesion d = 0.88; grammar mistakes reduced d = −1.12) and immediately after use (examiner total d = 1.12; cohesion d = 0.49), while also improving the subjective writing experience (lower Cognitive Load Theory d = −1.59; greater perceived ease d = 1.04). Notably, while students' sense of ownership dropped during ChatGPT use (d = −0.52), their perceived ownership increased after writing again unaided, with the "substantial contributions" item improving significantly (d = 0.64, p = .004), suggesting a nuanced authorship–ownership relationship.

Key Findings

Writing quality gains. ChatGPT statistically significantly improved students' writing quality both during use and immediately after. Examiner-rated total text scores (two teachers scoring five dimensions: word choice, language/grammar, scope, logic in outline, argumentative cohesion) rose from d = 1.18 (p < .001) during use to d = 1.12 after use — a gain of nearly one standard deviation in under an hour of assisted practice — with significant improvements in precision of word choice (d = 1.11), language and grammar (d = 1.49), and argumentative cohesion (d = 0.53), while scope and logic in outline showed only descriptive, non-significant differences.

Quality indicators and readability. Computed indicators corroborated the human ratings: readability (six ARTE-toolbox scores plus Grammarly and Wordcounter measures, α = .95) improved (d = 1.09 then 0.49), cohesion measured by TAACO (within-paragraph, α = .80) rose (d = 0.88 then 0.49) alongside givenness (d = 0.97 then 0.64), and grammar mistakes detected by Microsoft Word fell (d = −1.12 then −0.73); spelling dropped by a factor of five during use but was not significant.

Transfer to writing without AI. Benefits persisted when students wrote again without AI assistance, suggesting skill acquisition rather than mere output dependence — a tentative H2 finding, though the small ad hoc sample (n = 21) and single three-hour session warrant caution in generalizing to longer-term skill development.

Improved writing experience. ChatGPT improved the subjective writing experience — perceived ease (d = 1.04, p < .001), felt success (d = 0.69), and reduced cognitive load (d = −1.59) during use, with ease (d = 0.60) and success (d = 0.50) still higher afterward; topic understanding did not change, and perceived control was reduced during use (d = −0.56) but unchanged afterward. Students wrote on three permuted fashion-related topics (self-image in virtual avatars, AI's potential in fashion design, and VR embodiment/metaverse in the industry), drawing only on three preselected sources per topic to avoid confounding writing with literature review.

Authorship vs. ownership. While perceived ownership was significantly lower during ChatGPT use (d = −0.52, p = .014), it rose afterward (d = 0.30, marginal), and students rated "AI was writing, I was assisting" (M = 3.43) higher than "I was writing, AI was assisting" (M = 2.76), asking ChatGPT an average of 4.76 questions per task, with about 28 prompts aimed at generating novel text from the sources. The item "I made substantial contributions to the text" improved significantly after ChatGPT (d = 0.64, p = .004) — a nuanced distinction relevant to the knowledge base's Learner Agency and Cognitive Offloading concepts.

Implication. In-class AI writing support can enhance learning and skill acquisition in Higher Education, but the authorship–ownership tension, the small sample, and risks such as Over-Reliance and diminished Critical Thinking require careful pedagogical design and structured guidance on verifying AI output.

What this means for practice

• Instructors. Pair in-class AI writing with an unaided follow-up task: quality gains persisted when students wrote again without ChatGPT (total text scores d = 1.12), which is the evidence that the AI supported skill and not only the product.
• Instructors. Name the authorship question during the task rather than after it — students rated "AI was writing, I was assisting" (M = 3.43) above "I was writing, AI was assisting" (M = 2.76), and perceived ownership dipped while ChatGPT was in use (d = −0.52).
• Faculty developers. Design short supervised writing cycles (students spent 35–50 minutes per text) around a fixed source set as the unit of AI-writing professional development, since a single three-hour session moved quality, experience, and ownership.
• Designers. Restrict the source pool for AI-assisted writing tasks — students worked from three preselected articles per topic — so tool use supports synthesis rather than substituting for the search and selection work.

Limitations

• Of 25 students who began the study, 4 were excluded for incomplete data, leaving n = 21 first-year international fashion-design students from a single academic writing course in Berlin — an ad hoc sample the authors describe as not too familiar with academic writing.
• All three writing tasks, including the "after ChatGPT" measure, took place inside one three-hour session, so the study shows immediate persistence, not durable or semester-long skill development.
• AI assistance came from a GPT-3.5 API embedded in the experiment interface, so the effect sizes may not extend to newer models.
• Students wrote introductions from three preselected articles per topic, so results may reflect a simplified source set rather than unaided academic writing ability.

Citation

Bauer, N., Fütterer, T., Brucker, B., & Gerjets, P. (2026). Leveraging ChatGPT in academic writing: ChatGPT enhances students' writing quality, writing experience, and ownership. Computers and Education Open, 100351. https://doi.org/10.1016/j.caeo.2026.100351