Responsible student use of generative AI in academic writing: a scoping review of boundaries, educational approaches, and evidence gaps

Synthesis: Tie, Gil Ortega and Zhao asked how the literature separates responsible from irresponsible student use of generative AI in academic writing, how those boundaries become teaching and guidance, and what evidence exists that the approaches work. Following PRISMA-ScR without registration, they searched five databases for English- and Spanish-language higher-education studies from January 2022 to June 2026 plus citation searching, and moved from 21,720 records to 31 studies. Charting produced 192 practice-level records from 29 studies and 122 approach-level records from 29, with 27 studies contributing to both datasets: 107 (55.7%) Responsible, 42 (21.9%) Contested and 43 (22.4%) Problematic, with 138 carrying conditions from which six principles were synthesized. The conclusion is modest: the principles organize the grounds for judgment, not a list of permitted uses; only 51 of 122 approaches were both implemented and evaluated, mostly through self-report; and no study was appraised for quality.

Key Findings

1. Five databases plus citation searching produced 21,720 records; after removing 6,503 duplicates and 38 retracted, 15,179 were screened at title and abstract and 3,003 in full text, yielding 31 studies.
2. Of the 192 practice records, 107 (55.7%) were categorized Responsible, 42 (21.9%) Contested and 43 (22.4%) Problematic, and 138 carried conditions or notes.
3. The 122 approach records divided into 69 pedagogical, 19 guidance, 12 assessment, 8 policy-linked and 14 other; 52 were proposed only, 19 implemented without evaluation and 51 implemented and evaluated.
4. The clearest effect estimate came from Bravo-Duarte et al. (2026): a small improvement in self-reported verification behavior against a control group, F(1, 345) = 2.80, p < .05 one-tailed, ηp2 = .01, with no significant differential reduction in trust in ChatGPT, F(1, 346) = 0.31, p = .29, ηp2 = .001.
5. Jointly reported student and staff judgments amounted to 16 records from three studies: 3 Responsible, 9 Contested and 4 Problematic, which is 8.3% of all coded records but 21.4% of Contested ones.

What was searched, and what the counts count

The protocol was prepared in advance but not registered. The search covered Scopus, Web of Science, ERIC, the British Education Index and Education Source from 1 January 2022 to 6 June 2026, with Spanish records added in January 2026. The unit of synthesis is the record, not the study, so a paper reporting many practices contributes many times: 5 studies supplied 62 of 192 practice records and 37 of 122 approach ones. Record frequencies describe the composition of the extracted material, not independent observations or population prevalence. Broad AI and ethics vocabulary made the search imprecise: 21,720 records for 31 studies. Included studies were recent: 77.4% appeared in 2025 or 2026.

Six principles, and why no practice is responsible by itself

From the 138 conditioned records the authors synthesized six principles: targeting learning objectives, preserving human agency, verifying outputs, transparency and accountability, protecting authorship and human voice, and privacy, consent and fairness. They are interpretive, agreed by discussion among the authors, and never validated as decision criteria.

The conceptual finding is that a practice changes category with the assignment it sits in. Editing and language polishing was Responsible in 21 of 22 records, while drafting was Responsible in 11, Contested in 9 and Problematic in 13. Among 95 distinct practice labels, 12 (12.6%) appeared in more than one category. The operative variable is the assessed capability, not the tool.

Where the chain breaks, and what the corpus cannot support

Only four studies define a boundary, design an intervention around it and test whether decisions aligned (Bilikozen, 2024; Chehak, 2025; Judson, 2025; Wan, 2025). Evaluation concentrates on confidence, understanding and stated intention; direct writing measures appear in only two studies, Fadhly (2026) with pre-post rubric scores and Schrepel (2026) with legal redrafting comparisons, and Carter's (2025) reported gains in clarity rested on instructor observation. Transparency, authorship and privacy are widely endorsed and rarely evaluated, so the corpus cannot say whether disclosure produces transparency or compliance. The Contested category is thin: its three declaration records all come from Wu et al. (2026) and its single editing record from Sullivan et al. (2024). Discipline-specific evidence is thin: 22 of 31 studies were mixed-disciplinary, and nothing follows students over time or across tasks.

None of this licenses an effectiveness claim. "Implemented and evaluated" means some evaluation was reported, often of a program rather than a component; no study was appraised for risk of bias; and the 16 joint records cannot carry a student-versus-staff comparison. Nor does it test whether two instructors applying the six principles would reach the same verdict, which is what a decision aid needs.

What this means for practice

Instructors come first: the central finding is about assignment design. State the intended learning outcome before deciding what use is acceptable, and ask students to explain whether the assistance supports or replaces the capability being assessed. The review's three proposed activities are sensible and untested: explaining disclosure choices, comparing AI-assisted and unassisted drafts, and working through privacy scenarios. Administrators should treat institutional guidance as a floor rather than a ceiling: it can require verification, disclosure and data protection but cannot say how much drafting help is acceptable in one assignment. Researchers should connect boundary, intervention and evaluation in one design, measure process evidence such as prompts and drafts, follow students over time, and test whether independent instructors applying the six principles agree.

Limitations

• The search excluded gray literature and preprints, captured Spanish studies only when indexed under English terms, and used broad AI and ethics vocabulary, so the evidence base is narrower than the topic.
• Full-text screening and initial coding were done by the lead reviewer; dual screening covered only the 250-record calibration samples and the Spanish records; the revision audit covered 119 practice and 79 approach records from 17 of 31 studies.
• The unit of synthesis is the record, so 5 studies supplied 62 of 192 practice records and percentages are shares of records, not studies; no critical appraisal was performed, so implementation status is not evidence quality.

Citation

Tie, J., Gil Ortega, M., & Zhao, D. (2026). Responsible student use of generative AI in academic writing: a scoping review of boundaries, educational approaches, and evidence gaps. Frontiers in Education, 11, 1998449.