England's university regulator clarifies which AI practices fit its quality conditions and publishes the study behind them
The Office for Students and Advance HE have put out a 48-page report, based on more than 300 students and staff, and a guide that translates its findings into registration conditions B1, B2 and B4.
The report is based on a consultation carried out between June and September 2026: 16 events, 64 responses to open questions and 22 interviews, with input from 244 staff members and 91 students at no fewer than 127 institutions. Its authors warn that participants signed up voluntarily and so may not represent the whole population, and that the conclusions are those of the OfS and Advance HE, not regulatory requirements.
Citing surveys by the Higher Education Policy Institute (HEPI), the report notes that 94% of students said in late 2025 that they had used generative AI in assessed work, up from 53% in late 2023. It records valuable uses, especially for students with disabilities or whose first language is not English, and one central risk: offloading reading, problem-solving or writing onto AI can bypass the productive struggle needed to learn.
The assessment chapter is the most forceful. It describes how many unsupervised tasks can be completed to a good grade with little involvement from the student, and how the traffic-light systems (red, amber, green) that many institutions use to classify permitted uses are hard to enforce unless the design of assessments changes. On detectors, it concludes that their output should be, at most, one piece of evidence among others; the guide asks that misconduct proceedings rely on a balanced set of evidence, without “placing disproportionate weight on the outputs of AI detection tools”. It defends secure assessments (in-person exams, presentations, oral exams, practical tasks) while acknowledging that they may be less authentic. It adds: “If it becomes possible to complete an assessment using AI without the knowledge and skills being tested, the assessment should be replaced.” As an example, the University of Bath moved in September 2026 to two “lanes”: a closed one, with no access to AI, and an open one, designed on the assumption that it will be used.
On literacy and strategy, the report finds that students' training is uneven, that staff training is often optional with low take-up, and that at many institutions approaches grow out of the initiative of individual teachers. The University of Kent, as a case study, offers a ChatGPT Edu platform with compulsory prior training in literacy, responsible use and data protection, although using it is optional.
The guide is the regulatory part. It gives examples of practices likely to meet conditions B1 (academic experience), B2 (support) and B4 (assessment): reviewing the quality of materials and feedback produced with AI, telling students clearly which uses are permitted, and mitigating differences in access. It makes clear that it will not introduce new rules. In the future quality system it proposes, it expects institutions to show that their assessments stand up to AI, with a possible impact on their rating if there are doubts.
Among the documents in the library, it is related to the MIT committee report, which also calls for redesigning assessment; here the source is a regulator and the focus is regulatory compliance. Limits: it is a qualitative, self-selected study centred on England. For a teacher at any level, the most useful parts are chapter 2, with its redesign options, and paragraphs 53 to 55 on usage policies. To get started, the web guide is enough, and it can be read in a few minutes.
England's university regulator clarifies which AI practices fit its quality conditions and publishes the study behind them
The Office for Students and Advance HE have put out a 48-page report, based on more than 300 students and staff, and a guide that translates its findings into registration conditions B1, B2 and B4.
The report is based on a consultation carried out between June and September 2026: 16 events, 64 responses to open questions and 22 interviews, with input from 244 staff members and 91 students at no fewer than 127 institutions. Its authors warn that participants signed up voluntarily and so may not represent the whole population, and that the conclusions are those of the OfS and Advance HE, not regulatory requirements.
Citing surveys by the Higher Education Policy Institute (HEPI), the report notes that 94% of students said in late 2025 that they had used generative AI in assessed work, up from 53% in late 2023. It records valuable uses, especially for students with disabilities or whose first language is not English, and one central risk: offloading reading, problem-solving or writing onto AI can bypass the productive struggle needed to learn.
The assessment chapter is the most forceful. It describes how many unsupervised tasks can be completed to a good grade with little involvement from the student, and how the traffic-light systems (red, amber, green) that many institutions use to classify permitted uses are hard to enforce unless the design of assessments changes. On detectors, it concludes that their output should be, at most, one piece of evidence among others; the guide asks that misconduct proceedings rely on a balanced set of evidence, without “placing disproportionate weight on the outputs of AI detection tools”. It defends secure assessments (in-person exams, presentations, oral exams, practical tasks) while acknowledging that they may be less authentic. It adds: “If it becomes possible to complete an assessment using AI without the knowledge and skills being tested, the assessment should be replaced.” As an example, the University of Bath moved in September 2026 to two “lanes”: a closed one, with no access to AI, and an open one, designed on the assumption that it will be used.
On literacy and strategy, the report finds that students' training is uneven, that staff training is often optional with low take-up, and that at many institutions approaches grow out of the initiative of individual teachers. The University of Kent, as a case study, offers a ChatGPT Edu platform with compulsory prior training in literacy, responsible use and data protection, although using it is optional.
The guide is the regulatory part. It gives examples of practices likely to meet conditions B1 (academic experience), B2 (support) and B4 (assessment): reviewing the quality of materials and feedback produced with AI, telling students clearly which uses are permitted, and mitigating differences in access. It makes clear that it will not introduce new rules. In the future quality system it proposes, it expects institutions to show that their assessments stand up to AI, with a possible impact on their rating if there are doubts.
Among the documents in the library, it is related to the MIT committee report, which also calls for redesigning assessment; here the source is a regulator and the focus is regulatory compliance. Limits: it is a qualitative, self-selected study centred on England. For a teacher at any level, the most useful parts are chapter 2, with its redesign options, and paragraphs 53 to 55 on usage policies. To get started, the web guide is enough, and it can be read in a few minutes.
