Research Article
Making AI Work for Students: Reflections on Institutional Practice
Synthesis: A joint Office for Students and Advance HE study of 244 staff and 91 students from at least 127 institutions finds generative AI routine in student study and uneven in institutional response: unsupervised assessments no longer assure that students demonstrate what they are credited with, AI detection cannot carry an integrity case, and AI literacy provision is patchy while strategic adoption remains the exception.
Key Findings
- Students use AI for the parts of study that used to require the most effort. Explaining difficult concepts, summarizing long documents and generating practice questions are the common uses, and neurodivergent, disabled and English-as-an-additional-language students report the largest barrier reductions.
- Staff see learning atrophy. Participants described over-reliance on summaries, AI search replacing judgment about which text matters, and offloading displacing the productive struggle that develops understanding, especially in the early years of a degree.
- Unsupervised assessment is no longer valid on its own terms. Essays and coursework submitted remotely can be completed to a high standard with limited student engagement, and students reported peers receiving comparable grades for work they could not explain.
- Detection cannot bear the weight institutions place on it. Research suggests detectors are unreliable, improvements are temporary, and humaniser tools erode them further; some institutions still use scores as misconduct evidence.
- Redesign is happening, unevenly. Institutions are stress-testing assessments with AI, adding secure formats (exams, vivas, presentations, practicals), staged process assessment, and tasks where AI use is required; the University of Bath moved to a two-lane model in September 2026.
- Strategy is the exception and regulation is contested. Most approaches came from individual staff rather than institutional strategy, and participants asked for clarity about existing conditions of registration rather than new ones.
Background and Method
The Office for Students, which regulates English higher education, and Advance HE ran the research between June and September 2026: 16 live events including seven in-person roundtables in London, Manchester and Bristol, four online roundtables, two workshops at the Advance HE Teaching and Learning Conference and three listening events on OfS non-regulatory visits, plus 64 open-ended online prompt responses and 22 interviews with institutions, sector bodies, technology companies and graduate employers. The sample is self-selecting, and the authors say so: it is not representative of students or staff as a whole. The report mixes primary findings with wider literature and four institutional case studies, and it deliberately excludes research and the longer-term future of higher education.
Teaching and Study Practice
Students described AI as a study aid rather than a substitute: explanation, summarization and practice questions dominate, and the students who benefit most are those the system already disadvantages least, including neurodivergent students using transcription and spelling support, disabled students converting notes to braille, and students whose first language is not English decoding marking criteria and academic language. Use is not universal. Some students avoid AI on principle or after a misconduct process, and staff described a stigma in which AI use implies minimal effort, which discourages transparency and open discussion. Staff uses vary as much as student uses, from generating practice questions and case studies to producing model answers, and Feedback is the contested case: students criticized AI-generated feedback as poor quality or undisclosed, while staff saw it as a starting point that frees time for contact. Most institutions provide tools, usually Microsoft Copilot Chat under existing Microsoft 365 licensing, with Oxford, the London School of Economics and Kent buying other models; participants worried that premium tiers widen an existing digital divide, and that staff entering student data into free tools exposes records to training use.
Assessment
The report is blunt that many current assessments no longer measure what they claim: remote essays are described as particularly vulnerable, and staff told the researchers some assessments are not fit for purpose. Detection research is summarized sceptically. Human judgment of AI text is unreliable, detector improvements are followed by model advances, adversarial rewriting is cheap, watermarking is unproven, and AI-related misconduct cases rose to the 2024-25 academic year. Turnitin told the researchers its detector's output should be one data point only, yet staff reported institutions using scores as supporting evidence and students reported no route to prove authorship once flagged. Redesign takes three forms: secure assessments that keep AI out, process-based assessment with checkpoints, and assessments where AI use is required, such as critiquing model output. Secure formats are resource-intensive, and closed-book exams may disadvantage disabled students; validity, not security alone, is the test the report applies.
AI Literacy and Strategy
The report defines AI literacy across four areas: technical, evaluative, practical and ethical. Provision is uneven. Students arrive with different prior experience, the gap widens through study, and many reported limited or no training; the approaches that exist range from optional study-skills modules and capability frameworks to AI-integrated curricula and subject-specific tool training, with students' unions running their own workshops. Staff literacy is the enabling condition, and training is mostly optional with low take-up; institutions that made it compulsory report better reach, and the most valued training is discipline-specific. The University of Kent's ChatGPT Edu rollout in March 2026 makes training a precondition of access while keeping tool use optional. Strategy is less developed than practice: most respondents traced their institution's approach to individual experimentation, only a few could point to an AI strategy, and barriers included low confidence in outputs requiring heavy oversight, immature data infrastructure, slow approval processes and the risk of price rises. Institutional infrastructure and governance therefore matter more than another guidance document.
What this means for practice
- Administrators. Test every unsupervised assessment by attempting it with AI before the next cohort does; if it can be completed without the knowledge it claims to measure, replace it rather than relying on misconduct processes afterwards.
- Administrators. Treat detector output as one data point and never as the basis for a lowered mark; the report documents students with no route to disprove a flag, and Turnitin's own position agrees.
- Administrators. Fund the secure and process-based designs you adopt, including the staff time to build checkpoints and vivas, and check that closed-book formats do not disproportionately disadvantage disabled students.
- Administrators. Make subject-specific AI literacy part of the curriculum rather than an optional study-skills module, and require staff training before licensing tools, as Kent does.
- Administrators. Write a short strategy that sets principles and oversight, while leaving room for departmental experimentation, and remove the data-infrastructure and approval bottlenecks that stop pilots becoming services.
Limitations
- The sample is self-selecting: 244 staff and 91 students from 127 institutions chose to take part, so the findings describe engaged institutions rather than the sector as a whole.
- Fieldwork ran June to September 2026, a snapshot of capability and policy in a fast-moving area; several findings are explicitly hedged on detection research continuing to improve.
- The study reports perceptions and institutional practice, not learning outcomes: no student cohort was followed, and claims about atrophy and grade equivalence rest on participant reports.
- The four case studies are institutional self-descriptions, and the report excludes research uses of AI and longer-term questions about the future of higher education.
- The authors state the views are conclusions based on the research, not regulatory requirements set by the OfS, which limits how far the practice guidance can be read as compliance advice.
Citation
Office for Students & Advance HE. (2026). Making AI work for students: reflections on institutional practice. OfS 2026.56.