A student’s essay comes back with a detector score attached, and someone has to decide what that score means. A new research report from the Office for Students (OfS) and Advance HE offers evidence on how reliable that score is and on who is trained to act on it.
What the report found about detection and assessment
The report, Making AI work for students: reflections on institutional practice, is dated 11 November 2026. It says the research ran between June and September 2026. It draws on conversations with over 300 higher education students and staff, from over 120 institutions. The research explored primarily English higher education institutions’ responses to AI. The report presents findings and a set of prompts for institutions to reflect on. Participants generally suggested that new conditions of registration from the OfS on AI could be counterproductive. For that reason this column treats the report as evidence, and the actions at the end as practitioner suggestions.
On detection, the report is direct. It says evidence suggests that AI detection tools “do not consistently and reliably identify AI use.” On policy, it says students are frequently unclear about whether and how they are permitted to use AI in assessed work. It also says it is usually not possible to conclusively establish whether a student has adhered to their institution’s AI policy. On assessment design, the report says that where assessments are unsupervised, it is frequently possible to complete them to a high standard using AI tools without showing the knowledge and skills they are meant to test.
A detector score is Responsible AI, and a ruling is AI Governance
Responsible AI is machine checking machine. A detector that scores an essay fits that description: software checking work that may have been produced by software, with no person answering for the result. On the report’s own reading of the evidence, that check does not reliably identify AI use.
AI Governance closes the gap. Under Checkpoint-Based Governance (CBG), a named human decides accept, modify, or reject on the case and records why. In a university that person is an assessor or module lead. The detector score becomes one input to that decision. It cannot be the decision.
Who staffs the checkpoint
A named assessor can carry that decision only with the literacy to make it. The report defines AI literacy for students as “the knowledge, skills and capabilities students need to be effective, responsible users of AI” while they study and afterwards. It then says staff AI literacy is an essential enabler of effective institutional responses to AI. It adds that the most valuable training will often be subject-specific, because AI has different implications and applications across disciplines. The same passage reports that at present, AI literacy training for staff is frequently optional and often has low take-up.
Students face a similar gap. The report says the provision of AI literacy training for students is uneven, and that many students reported receiving limited training or none at all. Put the findings side by side. A student may not know the rule. The staff member judging the case may not have been trained to apply it. A tool score the report calls unreliable sits between them.
Institutional structure does not close the gap yet. The report says some institutions have appointed AI leads or ‘champions’ to help support staff AI literacy. It also says relatively few contributing institutions have adopted strategic approaches to AI and embedded these throughout their institutions. In its strategy findings it lists lack of confidence in AI outputs among the barriers to systematic integration, which means extensive human oversight is required. That oversight has to come from people who know what they are overseeing.
The measurement side of this problem is the subject of Measuring the Governance Competence at the Center of AI Literacy, which treats governance competence as the ability to direct, challenge, verify, and own work done with AI. An assessor who signs an AI-misuse finding is exercising exactly that competence, and optional training leaves it untested.
The report does not point only one way. It says secure approaches designed to prevent students from using AI will often be the most effective. It also says innovative approaches that responsibly integrate AI will often be positive, but need thoughtful design. Some participants called for greater regulatory clarity on complying with existing conditions of registration in the context of AI. A governance checkpoint works under every one of those assessment designs, because each one still ends with a person deciding a case.
Practitioner suggestions: who signs, what record, what measures
These are this column’s suggestions, not requirements from the report.
- Who signs. One named assessor per module signs any AI-misuse finding. A detector score alone never closes a case.
- What record. An accept, modify, or reject log for each case. Each entry states the evidence beyond the tool, such as drafts, version history, a conversation with the student, or a check of the cited sources, plus the reason in a sentence, the date, and the assessor’s name.
- What training. Assessors complete subject-specific AI literacy training before they sign findings.
- What students see. Every module carries a written statement of permitted AI use.
Four measures show whether the checkpoint works:
- Share of AI-misuse findings that rest on a tool score alone. The target is zero.
- Share of assessors who completed subject-specific AI literacy training.
- Share of modules with a written statement of permitted AI use.
- Appeal overturn rate on AI-misuse findings.
The first number to pull is how many findings last year rested on a score alone.
Sources
- Office for Students, & Advance HE. (2026, November 11). Making AI work for students: reflections on institutional practice. https://www.officeforstudents.org.uk/publications/making-ai-work-for-students-reflections-on-institutional-practice/
FAQ
Does the report set new rules for universities?
The page presents research findings and a set of prompts for institutions to reflect on. It also records that participants generally suggested new OfS conditions of registration on AI could be counterproductive. The suggestions at the end of this column come from the column, not from the report.
Can AI detection tools reliably identify AI use?
The report says evidence suggests they do not consistently and reliably identify AI use. A score from such a tool is an input to a decision, and a named person should decide the case.
What does the report mean by AI literacy?
For students, the report describes it as the knowledge, skills and capabilities they need to be effective, responsible users of AI while they study and in their future professional and personal lives.
Why does staff AI literacy matter for misconduct cases?
The report calls staff AI literacy an essential enabler of effective institutional responses to AI, and says staff training is frequently optional with low take-up. An assessor who rules on AI use needs training in how AI works in their own subject.
What is the difference between Responsible AI and AI Governance in this case?
Responsible AI is machine checking machine, such as a detector scoring an essay. AI Governance puts a named human at a checkpoint to decide accept, modify, or reject and to record the reason.
Which measures show the checkpoint works?
Track the share of findings resting on a tool score alone, the share of assessors who completed subject-specific AI literacy training, the share of modules with a written statement of permitted AI use, and the appeal overturn rate.
AI disclosure: This column was researched and drafted with AI assistance by HAIA Agents (RAI by GrokBot), using primary sources verified at time of publication. #AIgenerated
Leave a Reply