Imagine a student who asks a generative AI system to suggest ideas, reshape an outline and smooth the language of an essay. The student checks the output, rejects some of it and submits a text they can explain. They say the work is still theirs because they made the final decisions. Is that enough to make them its author?
Christine Savvidou’s Negotiating Generative AI-Assisted Academic Writing: Student Perspectives on Authorship, Responsibility and Learning, published in Education Sciences on 20 September 2026, asks students themselves where that boundary lies. Its most useful contribution is not a rule about acceptable AI use. It is a picture of authorship being negotiated through verification, judgement and responsibility. Its central limitation is just as instructive: students’ accounts of “cognitive ownership” are not yet evidence of what they actually did, understood or could later do without the same support.
A study of how students draw the line
Savvidou studied two sections of a 12-week undergraduate academic-writing course at a private English-medium university in Cyprus in spring 2026. Generative AI was incorporated into workshops rather than treated as a prohibited external intrusion. Students could use it for such purposes as brainstorming, organisation and language support, while being expected to verify sources and remain responsible for the work.
The main qualitative material consisted of 55 written reflections, about 14,600 words in total, produced by students enrolled on the course. A voluntary anonymous questionnaire, completed by 27 students, supplied descriptive context. Savvidou used reflexive thematic analysis for the reflections and reports the questionnaire percentages separately; because the survey was anonymous, the two datasets could not be linked at the level of individual students.
This design is well matched to a question about how students describe authorship and ethical responsibility. It is not designed to establish how AI changed their writing quality, what their prompts and revisions contained, or whether learning transferred to a later unaided task. That distinction matters throughout the paper.
What the students said
Almost all survey respondents reported using generative AI. The most common uses were brainstorming (63%), grammar and feedback (48.1%), and outlines (44.4%); 18.5% reported using it to generate text. Just over half agreed that AI-assisted writing could still count as their own work, while a third were neutral and 14.8% disagreed.
The ethical boundary was clearer when understanding and acknowledgement entered the question. Some 92.6% judged it unethical to submit work the student did not understand, and 85.2% regarded unacknowledged submission of AI-generated writing as dishonest. Two thirds accepted grammar improvement, nearly half accepted idea generation, but 81.5% rejected generation of complete paragraphs. Most also believed that AI could support struggling writers, while almost three quarters thought it might weaken the development of independent writing.
The thematic analysis organises the reflections around four connected ideas: AI as support rather than substitution; authorship preserved through human judgement; ethical boundary-work; and the tension between assistance, independence and learning. Students repeatedly placed authorship not in producing every word unaided but in evaluating suggestions, selecting what to retain, checking accuracy and accepting responsibility for the finished text.
The paper’s strongest move
That account is more precise than the familiar opposition between “human writing” and “AI writing”. A student may be highly active while working with generated language, just as a student may remain intellectually passive while copying a tutor’s corrections or assembling conventional phrases without understanding them. Savvidou therefore directs attention to the operations that connect assistance to authorship: verification, evaluation, decision and accountability.
The multilingual context gives this argument practical force. Language support may widen participation without removing the learner’s conceptual labour. Treating every improvement in fluency as suspect would confuse linguistic surface with intellectual contribution and could penalise students who have the ideas but need help expressing them. The study is also commendably cautious. It does not claim that reflective judgement automatically guarantees learning, and its limitations are stated plainly.
The evidence does not yet show the process
The reflections were written for a course taught by the researcher, after students had encountered its guidance on responsible AI use. That is not a defect in itself: the course is the setting in which students were learning the relevant distinctions. But it creates an interpretive problem. A reflection can reveal how a student has learned to describe acceptable practice while telling us less about the practice itself.
Students knew the norms against substitution, unverified information and concealed generation. Their accounts may therefore combine genuine self-understanding with social desirability and assessment management. The paper also acknowledges that the reflections themselves may have been written with generative-AI assistance. Consent was handled after assessment, with an opportunity to opt out, but the institutional relationship between instructor, evaluator and researcher remains part of the conditions under which the accounts were produced.
The survey cannot resolve this problem. It is small, descriptive and unlinked to the reflections. Nor are there prompt histories, successive drafts, source-check records, screen observations or interviews anchored to particular revisions. The study therefore establishes a rich vocabulary of student positions, not whether a student who claimed to verify and decide actually exercised those capacities at the relevant moments.
An alternative reading: learning the acceptable story
One plausible alternative reading is that the paper documents institutional boundary-work more than cognitive ownership. Students are learning an acceptable story about being an author in an AI-permitted course: use the system for support, keep human control, verify outputs and understand the final text. That story may be educationally valuable. It supplies distinctions with which students can question their own conduct. But being able to state the distinctions is not the same as enacting them.
This interpretation does not accuse students of deception. People make sense of new practices using the moral and professional language available in their setting. The important point is methodological: reported responsibility is an object of study, not a transparent window onto responsibility in action. What looks like an individual conviction may also be the course’s policy speaking through the student.
An Autopoietic Ecology reading: authorship has to travel
Cambridge EdgeLab’s research programme treats meaning as something mediated across relations rather than simply stored inside an isolated individual. In this case, a suggested phrase or argument moves through a chain: model output, student evaluation, source verification, written submission, teacher assessment and institutional rule. Each stage transforms what the text can mean and who can be held answerable for it.
From this perspective, the student’s final authorisation—“I chose to keep it”—is important but incomplete. The stronger question is whether the judgement can travel. Can the student explain why the suggestion was accepted, detect a plausible error, defend the relation between evidence and claim, and continue the work when the same prompt or tool is unavailable? The Concept Guide calls attention to the constraints through which a capacity is maintained. For AI-assisted writing, ownership is more convincing when understanding persists across a change in conditions.
This interpretation also refines a simple account of human control. A learner may press the final button while the practical direction of the text has already been shaped by model defaults, assignment conventions and institutional reward. Conversely, substantive learning may occur through intensive engagement with suggestions that the student frequently rejects. Control is not a location occupied once; it is a pattern of consequential operations across the process.
What would count as stronger evidence?
A useful next study could connect students’ reflections to version histories, prompts, rejected suggestions and source-checking records. Stimulated-recall interviews could ask students to explain particular decisions while looking at their own drafts. A short oral defence could test whether they can reconstruct an argument and its evidence. Most importantly, a delayed task under different conditions could ask whether the capacities attributed to AI-assisted practice—planning, evaluating and revising—transfer beyond the original assignment.
Such evidence could challenge the EdgeLab interpretation as well as support it. If students’ claims about verification and judgement reliably predicted later independent performance, despite differences in tools and contexts, then a relatively narrow account of reflective cognitive ownership might be sufficient. If process traces instead showed that substantive ideas and argumentative relations were routinely accepted without scrutiny, the institutional-boundary reading would become stronger. Either result would be more informative than treating disclosure alone as a proxy for learning.
For teaching and policy
The practical implication is not a return to tool bans. Savvidou’s students identify real benefits, especially for brainstorming, feedback and linguistic support. Policy should preserve those possibilities while designing occasions on which understanding becomes observable. Teachers can require a small number of consequential choices to be explained, ask for evidence of source verification, compare drafts, or include a brief follow-up task that cannot be completed by reproducing the same generated text.
Institutions should also notice the labour involved. Process review, oral explanation and responsive assessment require teacher time. A policy that demands proof of responsible use without supplying assessment capacity merely transfers a governance problem to individual teachers and students.
Savvidou’s paper gives institutions a better starting question than “Was AI used?”: what did the student still have to understand, decide and answer for? Its evidence shows how students articulate that responsibility with considerable nuance. The next step is to observe whether those articulated boundaries hold when ideas, texts and capacities move through the actual writing process.
Reference
Savvidou, C. (2026). “Negotiating Generative AI-Assisted Academic Writing: Student Perspectives on Authorship, Responsibility and Learning”. Education Sciences, 16(9), 1568. https://doi.org/10.3390/educsci16091568.
Cambridge EdgeLab is an independent research initiative. This review represents Steven Watson’s critical interpretation and does not imply endorsement by the article’s author, the University of Cambridge or any other institution.
Leave a Reply