Scoping review maps six principles for responsible AI use in student writing

A 2026 scoping review of generative artificial intelligence in higher-education writing concludes that students’ use of tools such as ChatGPT cannot be labelled responsible or irresponsible from the tool alone. Across 31 studies, the strongest recurring distinction was whether AI assistance supported the learning objective of an assignment or replaced the intellectual contribution students were meant to make. The researchers synthesised their findings into six principles: targeting learning objectives, preserving human agency, verifying AI output, ensuring transparency and accountability, protecting authorship and human voice, and addressing privacy, consent and fairness. The result is a descriptive map of current practice rather than evidence that any particular classroom intervention is effective.

The review followed the PRISMA-ScR reporting guideline and searched Scopus, Web of Science, ERIC, the British Education Index and Education Source for English- and Spanish-language higher-education studies published from 1 January 2022 to 6 June 2026. Citation searching was added, and studies qualified if they supplied original empirical data or systematically documented, first-hand evidence of a practice already enacted with students. Twenty-eight studies met the empirical-data criterion and three entered through the practice-based route. Searches produced 21,720 records; after duplicate and retraction checks, 15,179 records were screened at title-and-abstract level and 3,003 at full-text level. The final corpus contained 31 studies, all published between 2023 and 2026, with 77.4% appearing in 2025 or 2026. Most concerned mixed-disciplinary settings, undergraduate or postgraduate coursework, or other academic-writing contexts.

From 29 contributing studies, the researchers charted 192 practice-level records describing particular uses of generative AI. Of these, 107, or 55.7%, were classified as responsible, 42 as contested and 43 as problematic. These proportions describe the extracted material and should not be interpreted as population prevalence, because studies that documented more practices contributed more records. The classification depended strongly on where AI entered the writing process. Language polishing and editing were generally treated as responsible when students retained control and evaluated suggestions. Brainstorming could also be acceptable when its output served as a prompt for further thought, but idea generation became problematic when formulating the research question was itself the assessed skill. AI-generated references and summaries were considered problematic if students treated them as authoritative evidence without checking the original publications, while using generated material as a draft or model was often contested because acceptability depended on how much of the student’s contribution remained identifiable.

Three of the six principles concerned decisions made directly during writing. The first, targeting learning objectives, asks whether assistance supports or displaces the capability an assignment is intended to develop. The same tool could therefore be acceptable in an advanced course but not in a beginners’ writing class. The second principle, preserving human agency, distinguishes a scaffold from a replacement. Responsible uses exposed students to possible ideas, revisions or feedback, after which they selected, rejected, adapted and explained the material. Contested cases involved uncertain participation, while problematic cases delegated substantive reasoning or final judgement to the system. The third principle requires verification. Definitions, literature suggestions, quotations, citations and factual claims generated by AI were treated as starting points only when students located and checked the originals. Because hallucinations, bias and tool limitations remain possible, students remain responsible for the accuracy and integrity of the final text.

The remaining principles concern accountability and protection. Transparency and accountability meant acknowledging assistance and following relevant course, institutional or journal requirements, although disagreement remained over whether minor uses such as grammar correction or limited brainstorming required disclosure. Protecting authorship and human voice focuses on ownership, originality and personal expression, especially in multilingual writing: AI-supported translation or polishing may reduce linguistic barriers, yet it can also displace culturally significant features or standardise a student’s voice. Privacy, consent and fairness address the conditions of use rather than the text alone, including confidential manuscripts, sensitive information, undisclosed AI evaluation of student work, unequal access to paid platforms and biased outputs. Comparisons involving 16 jointly reported student and staff records found agreement that grammar correction and paraphrasing one’s own text could be responsible, while full-text generation and replacement of human judgement were problematic. Divergence clustered around the uncertain boundary between support and over-reliance.

The review also mapped 122 educational approach records from 29 studies. Pedagogical approaches were the largest group, with 69 records, followed by guidance with 19, assessment approaches with 12 and policy-linked guidance with eight; 14 concerned supporting measures such as infrastructure, monitoring, accessible resources and modelling. Common teaching methods included workshops, induction sessions, course-embedded AI literacy, guided practice and reflection. To operationalise verification, educators asked students to triangulate AI-suggested references through conventional searching, check facts and sources, and compare their summaries with AI-generated alternatives. To strengthen human agency, some instructors asked students to request lists of suggested corrections and apply the changes themselves, while others used generated outlines, feedback and images as objects for evaluation rather than finished prose. Approaches focused on learning objectives by defining permitted and prohibited uses separately for each writing task.

Implementation was mixed. Of the 122 approach records, 52 were proposed without reported implementation, 19 had been implemented without evaluation and 51 had been implemented with some form of evaluation. The evaluated group included 36 pedagogical approaches, nine guidance or policy-linked approaches, three assessment approaches and three supporting measures. Assessment redesign proposed making the writing process visible through planning notes, version histories, prompt records or reflective accounts, and pairing written work with oral examination. Module-level guidance was often tied to learning outcomes and student developmental level, while institutional proposals included collaborative policy development, staff training, regular review and provision of a common platform intended to support equal access. The review notes, however, that an evaluation of an overall programme does not establish that every component or every principle within it was tested.

Evaluations were strongest for proximal outcomes such as students’ confidence, understanding of institutional expectations, awareness of AI limitations, prompting and reported verification behaviour. One study reported a small but statistically significant improvement in self-reported verification relative to a control group, with a partial eta squared of 0.01 and p below 0.05, but the associated reduction in trust in ChatGPT was not significant. Other studies used pre-post self-reports, post-class reflections, classroom artefacts and instructor observations. Direct evidence on writing performance was less common, appearing in rubric-based pre-post writing assessments and comparisons of legal-redrafting across classroom conditions. Qualitative findings indicated that students could identify weaknesses in AI prose and articulate context-specific boundaries. These results show that instruction can accompany changes in reported understanding and practice, but they do not establish a general effect on writing quality, academic integrity or long-term responsible use.

The clearest research gap is the missing chain between a stated boundary, an intervention designed around it and an evaluation of whether students follow it. Although 27 of the 31 studies contributed to both the practice and approach datasets, contributing records did not mean those elements had been linked in a tested programme. The synthesis therefore supports a division of responsibility: institutions could establish baseline expectations for verification, disclosure, accountability and data protection, while individual courses specify how AI assistance relates to assessed learning. Teaching designs could make prompts, AI outputs, successive drafts and revision decisions visible, allowing instructors to discuss whether assistance supports or replaces the intended skill. The authors also call for discipline-specific work, because evidence from mixed settings cannot show how a boundary changes between subjects, academic levels or assessments. Future studies should test decision-support tools derived from the six principles and measure disclosure accuracy, source checking, retained authorship and transfer across new tasks.

The review has important limitations. It excluded grey literature, preprints and conceptual or policy-only work, and Spanish-language retrieval depended on the English search terminology. Full-text screening and initial coding were primarily conducted by one reviewer, although co-authors later audited 119 original practice records and 79 approach records from 17 studies. No formal risk-of-bias appraisal was conducted, and the six principles emerged through interpretive synthesis and author discussion rather than a validated classification system. Record-level counting also gave greater numerical weight to studies with more detailed reporting. With those constraints in mind, the safest conclusion is not that generative AI has a proven place in academic writing, but that responsible use cannot be specified as a universal list of tools or prohibited functions. Its defensibility depends on the assignment’s purpose, the student’s intellectual contribution, the safeguards attached to use and whether responsibility remains with a person who can verify, explain and defend the final work.

Subject of Research: Responsible generative AI use in higher-education academic writing

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

Article References: 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, Article 1998449. https://doi.org/10.3389/feduc.2026.1998449

Image Credits: AI Generated

DOI: 10.3389/feduc.2026.1998449

Keywords: generative AI, academic writing, higher education, student learning, academic integrity, AI literacy, human agency, output verification, scoping review, assessment design, Responsible, student