Generative AI in assessment has created an uncomfortable problem. A student can evaluate an AI response, edit it, and still misunderstand how much of the final work is genuinely theirs. The polished product may conceal a weak judgment about authorship, knowledge, and learning.
Jack Walton and colleagues examine this problem by returning to students’ actual AI interactions. Their study belongs within the wider discussion of AI and assessment because it asks what happens inside the process of completing a task, where many consequential decisions remain invisible to the teacher.
Summary of the Study
The researchers interviewed 26 undergraduate students from different disciplines at a large Australian university. Participants identified themselves as motivated users of generative AI. Most interviews used a “scroll-back” method: students opened earlier AI conversations, assessment drafts, or other digital traces and walked the interviewer through what they had done and why.
The team then analysed assessment-related episodes as judgment events, looking at the connected choices, actions, tools, purposes, and meanings surrounding each decision.
Six kinds of judgment appeared. Students used AI to acquire content knowledge or learn how to perform a process; discovered the tool’s limitations through repeated use; relied on it for tasks they felt unable to complete alone; adopted ideas with little critical appraisal; misjudged AI contributions as their own; and sometimes submitted AI-generated material without evaluating it.
These categories complicate simple divisions between legitimate assistance and cheating. The same tool could support learning in one episode and bypass it in another, depending on what the student knew, what the task demanded, and how the output was handled.
The most troubling cases involved self-judgment. Some students believed they had substantially altered AI-generated text, while comparison with their saved interactions and final assignments showed little change. Others treated novelty as evidence of quality or trusted an answer because they lacked enough subject knowledge to identify what was wrong.
This extends the idea of evaluative judgement in the AI age. Students must evaluate AI output, but they also need to judge their own knowledge, dependence, and contribution. Those judgments become especially fragile when fluent output creates an impression of understanding.
The study’s scroll-back design is a real strength because it does not rely entirely on general recollections. Even so, students chose which traces to bring, some reconstructed interactions during the interview, and the researchers’ narrative analysis was deliberately interpretive.
The sample was self-selected from one university and included motivated AI users, so the six categories map possibilities; they do not tell us how common each behaviour is. The study also did not measure learning outcomes. Its evidence supports careful claims about judgment processes, not causal claims about whether AI improved or damaged learning.
Practical Insights
Assessment design needs to make the student’s contribution easier to see and easier for the student to examine.
- Ask students to identify their intellectual contribution. A short commentary can explain which claims, interpretations, decisions, and revisions came from the student. Asking only what AI contributed leaves the student’s own role undefined.
- Use selected process evidence. Students might submit an early plan, one important AI exchange, a marked revision, and a brief explanation of what changed. This provides more insight than an enormous transcript or a generic AI-use declaration.
- Test knowledge and process together. AI often delivers a correct-looking answer while hiding how that answer was reached. Tasks can require students to explain the reasoning, apply it to a new case, or defend why a method fits the situation.
- Build checkpoints before submission. Students can compare a draft with the criteria, identify where AI influenced it, and complete a short independent reconstruction. These pauses may help counter the unexamined reliance associated with metacognitive laziness.
- Remove easy routes around the intended learning. If an online quiz can be completed by pasting questions into a chatbot, warnings about misconduct will have limited force. This is fundamentally a problem of assessment validity: the task must still generate credible evidence of the learning it claims to assess.
- Treat reliance with care. Some students used AI to manage writing difficulties, disability, anxiety, or gaps in confidence. Teachers need to examine what support makes participation possible and which parts of the task must remain with the learner. Accessibility and learning cannot be settled through a blanket rule.
For researchers, the scroll-back interview provides a useful way to place student explanations beside behavioural traces. Future studies could combine this method with version histories, independent performance tasks, and delayed assessment. That would help distinguish productive reliance that builds capability from reliance that produces acceptable work while leaving the learner unable to repeat it.
The paper leaves us with a question that should sit inside every AI-supported assessment: can students explain what they contributed and show the judgments behind it? If they cannot, the final submission may tell us very little about what they know or can do.
Reference
Walton, J., Bearman, M., Crawford, N., Tai, J., & Boud, D. (2026). How university students work on assessment tasks with generative artificial intelligence: Matters of judgement. Assessment & Evaluation in Higher Education, 51(6), 1066–1082. https://doi.org/10.1080/02602938.2025.2570328
