On this page

Synthesis: McDonald, Saunders, Kennedy and Tansell (2026) ran three focus groups with 19 students — 17 undergraduates and 2 taught postgraduates — across every department in the Faculty of Science at a Russell Group university, asking how assessment felt in a course where generative AI was prohibited unless a module said otherwise. The students were not asking for permission and not asking for a ban. They wanted assessment that certifies something a person did, and they read the technology through that lens: useful for finding a paper, useless as a substitute for understanding it, and corrosive where it replaced the skill a degree is supposed to certify. Three themes organize the analysis: human-led learning in which GenAI moves from threat to learning partner, equity problems that run in both directions, and the value students place on flexibility, diversity and authenticity. The authors read the findings through Universal Design for Learning and constructive alignment, and their recommendations ask institutions to license tools, publish criteria, diversify formats and treat students as co-designers rather than as suspects.

Key Findings

  • Students draw a sharp line between scaffolding and substitution. The approved pattern is using a tool to make a resource and then checking it, or using search summaries to locate a primary source and evaluating the source itself. The rejected pattern is handing over the task: one student described putting a PowerPoint in and receiving notes with nothing left to do, and said plainly that this is not the best way.
  • Deskilling is a discipline-specific fear, not a general anxiety. A computer science student said that before ChatGPT he would code and understand what he was doing, and that the tool now makes it easy to skip that. Another argued that professional contexts restore the requirement — live coding in interviews — so some assessments should certify unaided capability. Closed-book examinations were described as the only reliable way to guarantee no AI use, and then criticized as unfair because they test memory rather than ability.
  • Academic integrity is reframed as a question about who is being assessed. Rather than debating detection, students asked whether the work is theirs: "am I really answering the questions anymore? Are they now assessing me or the AI?" One described the shift in institutional guidance as disorienting, having been told for years that the tool was forbidden.
  • Trust depends on discipline fit and on bounded tools. Students trusted GenAI less in mathematics and physics, where errors are checkable and consequential, and more when a tool was purpose-built for the subject. The clearest example is a departmental AI notebook for a lab project that returned numbers but refused to write the report, leaving interpretation to the student; the same student noted that pasting the numbers into ChatGPT would have produced the entire report.
  • Equity concerns run in both directions. Requiring GenAI penalizes students who were never taught to use it and who lack the hardware; not using it while peers do feels like accepting a disadvantage. Paid versus free versions were named as a bias, and group work was identified as the place where unequal use turns into unequal workload.
  • Closed-book examinations are themselves an equity problem. Students described them as rewarding memory and compressing everything into a single high-stakes performance, with one asking whether a professional actually needs to recall everything. Coursework, remote open-book assessments and varied formats were preferred because they let students control the environment and demonstrate strengths.
  • Choice is valued, but only when options are equivalent and legible. Students liked choosing between a video and an essay, then reported that peers feared the essay would be marked worse. Choice without shared criteria and clear grading logic produces strategic guessing rather than inclusion.
  • Authenticity is the strongest signal of a worthwhile assessment. An OSCE and a group presentation in pharmacy were praised as rehearsal for professional practice; portfolios and presentations in mathematical sciences were described as useless to the student's learning. The difference students drew was relevance to the work the degree leads to.
  • Rubrics, formative drafts and criteria workshops build assessment literacy. Detailed rubrics that can be used as a checklist while writing, and the chance to submit a shortened draft for feedback, were named as the practices that made expectations visible.
  • The guidance students want is about best practice, not only restriction. One observed that embracing the technology is unavoidable because peers will use it regardless, and another asked for guidelines that describe good use as well as prohibited use.

How the study was done

The study took place in the Faculty of Science at the University of Nottingham during the 2024/25 academic year, as part of a larger project with Advance HE on inclusive learner experiences and belonging. Participants were 19 students from Biosciences, Chemistry, Computer Science, Mathematical Sciences, Pharmacy, Physics and Astronomy, Psychology, and Natural Sciences — 17 undergraduate and 2 taught postgraduate, aged 18 to 24 (M = 20.58, SD = 1.64). Recruitment ran through student course representatives and student communications, with an inconvenience allowance. Three focus groups met between March and April 2025 for 60 to 90 minutes each, were audio recorded and transcribed verbatim.

The protocol has an unusual feature worth naming. Focus group questions were co-created with two undergraduate student consultants from mathematical sciences and natural sciences, who were paid for the work and who also facilitated the sessions and transcribed the recordings. The authors' rationale is that peers elicit more open talk than staff, reducing the power gradient. Questions were deliberately left open rather than asking directly about links between inclusive assessment and GenAI, and the connections were made at analysis.

Analysis was inductive thematic analysis following Braun and Clarke, within an essentialist and realist frame, with themes developed from the data rather than imposed. The authors describe reflexivity work — revisiting the data and examining how their own interests in inclusive practice and GenAI shaped what they attended to. Ethical approval came from the School of Psychology ethics committee.

The institutional context matters for reading the findings. At the time, the university had no nuanced position on GenAI: use in formal assessment was prohibited unless specified, and suspected cases went through academic integrity processes, even as GenAI was being embedded in the reading, writing and coding tools students used daily. The authors present the study as a baseline taken before guidance matured.

What this means for practice

  • Publish what each assessment certifies, and do not outsource that judgment to a detector. The deskilling concern is answerable only if a program can say which outcomes require unaided performance and why; where a task must be done without tools, the justification should be the competence being certified, as in the live-coding example, not suspicion of students. The literature the authors cite finds that AI detectors misclassify non-native English writing, which makes detector-led enforcement an equity risk rather than a safeguard.
  • License the tools or do not require them. Students named paid versus free versions as a fairness problem, and one noted that not everyone has been taught to use a tool properly. Institutional licenses and a taught baseline are the minimum for any assessment that expects GenAI use.
  • Diversify formats and make the options equivalent. Multiple routes to demonstrate learning support students with different strengths, but only when criteria, rigor and marking logic are shared and explained. Without that, choice becomes a calculation about which route is marked generously.
  • Build assessment literacy in, and ask for evidence of process. Rubrics students can use while drafting, formative submission of a shortened version, and workshops on what the criteria mean were named as what made assessment legible. Asking for a short account of what a tool contributed, or a commentary on the AI-supported part of the work, makes disclosure meaningful rather than performative and creates the accountability moment that changes how students use the technology.
  • Co-design with students. The study's method is also its recommendation: paid student consultants shaped the questions, ran the groups and transcribed them, and the findings reflect what peers will say to peers.

Limitations

  • One institution and one faculty. All participants were science students at a single Russell Group university, so the discipline mix and the institutional context are fixed. The authors note that representation across schools within the faculty is a strength, not that the findings generalize.
  • Self-selecting, self-reported. Nineteen volunteers talked about their own practice; no observation of actual tool use, no logs, and no comparison with students who declined to take part.
  • The sample is unusual in composition. 57.9% reported a learning difference or disability, 84.2% were UK students, and the group skewed young (18 to 24). Those proportions are not typical of a faculty population and may shape the emphasis on flexibility and fairness.
  • The policy context has moved. Participants studied under a near-blanket prohibition at a point when critical AI literacy teaching was not widespread. The authors present this as a baseline, and caution that students now experiencing more nuanced guidance may report differently.
  • No outcome measures. The study reports perceptions, not learning gains, and cannot show whether the assessment designs students prefer produce better or more equitable results. Sector figures it cites — including 94% of students reporting GenAI use for assessed work — come from other surveys (Stephenson & Armstrong, 2026), not from this sample.
  • The link between inclusion and GenAI was inferred at analysis. Because questions were deliberately open, the connection between the two topics is the authors' interpretation of what participants said, not a topic participants were asked to address directly.

Citation

McDonald, S., Saunders, G., Kennedy, E., & Tansell, A. (2026). Inclusive assessment design in the age of GenAI: Advancing pedagogy and practice. Open Scholarship of Teaching and Learning, 5(2), 77–97.

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.