The conflicted student and the internal battle over generative AI use

Synthesis: Dewitt, Brumby, Bodei and Siromahov (2026) take the crisis narrative about students and generative AI literally and test its implicit premise — that students will cheat if they can. Two anonymous surveys of first-year psychology students at University College London, 94 quantitative and 97 qualitative respondents, describe a student body that mostly uses the tools for support, is well aware of their limits, and feels bad about it. The majority described an invisible line between support and doing the work, and many reported guilt at having crossed it. Alongside the self-justifications, the authors document a second layer through discourse analysis: students work rhetorically to preserve a sense of authorship while using the tools, which makes their accounts evidence about moral reasoning as much as about behavior.

Key Findings

1. Usage is common but not constant. Only 4.21% reported using genAI every day and 9.47% said they never use it, leaving the vast majority using it moderately.
2. Assignment help is the dominant use. 81.05% had used a genAI tool to help complete an assignment, which is the behavior the integrity debate centers on.
3. Students see it as a tool for support. The largest theme was "only a tool / just for support", with a substantial minority reporting it as a personal tutor for explaining concepts they were hesitant to ask about in class.
4. They are aware of the limitations. Most agreed the tools struggle with complex tasks, can produce out-of-context output, can be factually wrong and can be biased — the opposite of the naive-user picture.
5. A refusal minority is as large as the cheater minority. Some students decline to use genAI at all on ethical grounds, a group comparable in size to those who use it to do the work.
6. The costs are named by students themselves. Themes include cognitive debt, the loss of the learning opportunity, environmental impact, and social concerns about AI more broadly.
7. Continuing use is explained by pressure, not endorsement. Students named being disadvantaged relative to peers, temptation, time pressure and habit as reasons they keep using tools they have reservations about.

What the two surveys did

The quantitative instrument went to a cohort of 120 first-year BSc Psychology students, with 94 responses, and asked about frequency, uses, beliefs about learning support, awareness of limitations and the importance of authorial voice. The qualitative survey was open-ended and went to a combined cohort of 190, yielding 108 responses and 97 after eleven students withheld consent for publication.

Both were anonymous by design, requesting no demographics in order to maximize honesty, which is a defensible trade for a study about a norm violation and a real cost for interpretation. The authors describe the cohort's composition separately: 85.8% registered female, mean age 19.2, from university records rather than from the survey.

The invisible line

The phrase students reached for repeatedly is a line between using the tool in a supportive way and letting it do the work. The authors connect that line to the older literature on plagiarism, where students rephrase borrowed text just enough to bypass detection, and note the interactive nature of student-AI interaction makes the line easier to cross: the tool suggests, encourages further use, and each step looks small.

What makes the paper more than a survey is the second analysis. Following Billig's rhetorical approach, the authors read the responses as argumentative performances rather than neutral reports, identifying disclaimers, extreme-case formulations and preambles that let a student condemn cheating in general while justifying their own use. The guilt and the justification are both data: they show students holding a norm they believe they are partly violating.

Why the crisis framing fails

The crisis narrative assumes capacity implies action. The survey finds that most students want to learn, many value authorship highly — the authors report that most rated authorial voice as very or extremely important, with none saying it was not at all important — and a non-trivial group refuses outright.

That does not make the problem disappear. The students describe a genuine struggle, and the authors' point is that the struggle is the finding: a student who uses the tool while believing it may be damaging their learning is in a different situation from one who sees no issue, and needs a different response from an institution. Telling that student the rules will not resolve a conflict they already feel; changing what the assessment asks them to do might.

What this means for practice

• Instructors. Start from the conflict students already report. Most of this cohort knows the risks and feels the tension, so a lecture on integrity adds little; a task that makes the tool's role explicit and defensible does.
• Instructors. Ask about authorial voice directly. Students rated it highly and described rhetorical work to protect it, which suggests assignments that make the student's own contribution visible are aligned with what they already value.
• Administrators. Distinguish refusal, support and substitution when writing policy. The survey finds all three in one cohort, and a rule aimed at substitution will be read by the other two groups as an accusation.
• Researchers. Treat self-reports on integrity as rhetorical data. The discourse analysis shows students managing a norm while describing their behavior, so a survey that codes answers literally will overstate consistency.

Limitations

• Both surveys are from a single first-year cohort at one institution (University College London), and the qualitative sample is 97 students from that and one neighbouring program.
• The surveys were anonymous and collected no demographics, so the authors cannot link responses to individuals, prior attainment or discipline, and rely on cohort records for composition.
• Usage and beliefs are self-reported and cross-sectional, with no observation of actual behavior and no longitudinal follow-up.
• The quantitative instrument is descriptive: there is no control group, no intervention and no statistical model relating beliefs to use.
• The discourse analysis is interpretive, conducted by one author within a stated rhetorical framework, and the authors present the themes as an account rather than as a validated coding scheme.

Citation

Dewitt, S., Brumby, D., Bodei, S., & Siromahov, M. (2026). The conflicted student and the internal battle over generative AI use. PsyArXiv preprint.