England's OfS Tells Universities to Grade Students Without AI

England's Office for Students told universities on November 11, 2026, that they must ensure assessments measure what students know and can do, often by using secure, AI-free formats. The regulator warned that many unsupervised assignments can now be completed to a high standard with AI while requiring limited engagement from students.

The statement came alongside a joint research report with Advance HE and formal guidance for universities and colleges. The guidance applies the OfS's existing quality conditions to artificial intelligence rather than creating a new AI-specific registration condition. Those conditions require up-to-date courses and relevant skills, sufficient student support, and credible awards based on valid, reliable assessment. Institutions that fall short can face specific registration conditions, monetary penalties, or suspension of access to student support funding or OfS grant funding.

Josh Fleming, the OfS's director of strategy and delivery, described AI as both an opportunity and a "generational risk to the system as it stands, threatening to undermine students' learning and institutions' ability to assess them." He called for far greater use of AI-free assessment.

Why the regulator acted now

The stakes are straightforward. If AI assistance cannot be detected, the OfS argues, the value of an English degree depends on whether assessment design can still demonstrate what a student actually knows and can do. That question now sits inside an existing regulatory regime with real financial teeth: suspension of student support funding would hit institutions directly.

The report draws on research conducted between June and September 2026 with 244 higher education staff and 91 students or student representatives from at least 127 institutions. Evidence came from live events, 64 responses to online prompts, and 22 interviews. The OfS notes a methodological caveat: participants were self-selecting and may not represent wider student and staff populations.

What did students and staff report?

The evidence describes a gap between how assessment works on paper and how students complete it in practice.

• Students told researchers that peers sometimes completed assessments with large language models in just a few minutes and received grades similar to those earned independently.
• Some said those peers could not recall basic details of their own submissions.
• Staff reported high volumes of submissions sharing the same style and references.
• Remotely submitted essays were identified as especially vulnerable.

The pattern matters because the report's central regulatory conclusion follows directly from it. The report concludes that human reviewers cannot reliably identify AI-generated text and that detection tools do not consistently identify AI use. Consequently, policy declarations and retrospective misconduct proceedings cannot substitute for structurally sound assessment design.

The OfS sets limits on how detection evidence can be used. Detector outputs should be at most one piece of evidence in a fair process, and suspicion without strong evidence should not justify a lower mark. That position removes the most common institutional shortcut — running submissions through a detector and penalizing flagged students — and pushes the burden onto assessment design instead.

What counts as secure assessment?

The guidance defines secure assessment as a format designed to prevent AI use. The listed formats include:

• in-person exams
• presentations
• oral examinations
• performances
• practical or applied tasks

Where AI is permitted, students must still demonstrate the intended learning outcomes, which can include effective AI use. The guidance sets a replacement test: an assessment should be replaced if it can be completed with AI without the knowledge or skills it is meant to test.

One case study in the report illustrates the design choice. The University of Bath adopted a two-lane model in September 2026. The move followed a finding that the middle category in its previous traffic-light system, which allowed specified AI uses, could not be policed reliably and equitably. Under the new model, closed-lane assessments prevent access to AI, while open-lane assessments permit it and must be designed on that assumption. Both lanes still need to measure the intended learning outcomes.

Bath's experience is the concrete comparison for the sector: an institution that tried to regulate partial AI use, found enforcement impossible, and simplified to a binary model rather than a graded one.

What does secure assessment cost?

The report does not present AI-free assessment as free. Participants cited three main costs:

• the resources required for exams and oral tests
• concerns that closed-book exams may not reflect professional practice
• the possibility that some formats may put disabled students at a disadvantage

Institutions retain a legal duty not to discriminate and may need reasonable adjustments, including changes to assessment methods, while ensuring those adjustments do not create an unfair advantage. That balancing requirement makes secure assessment a design problem rather than a simple instruction to return to exam halls.

Is the OfS banning AI in education?

No. The report explicitly does not recommend removing AI from education. Students use it to explain difficult concepts, summarise documents, generate practice questions, and revise. The report identifies potential benefits for disabled and neurodivergent students and for those whose first language is not English.

The OfS draws a line between those activities and cognitive offloading, where AI replaces reading, research, problem-solving, or other demanding work. Offloading, the report argues, can bypass the productive struggle needed to develop knowledge and skills. The distinction — AI as study aid versus AI as substitute for learning — runs through the entire guidance.

Student demand points the same direction. Many students see AI literacy as important for employment and want more subject-specific training. Provision is uneven, and staff training is often voluntary with low take-up. Relatively few participating institutions had embedded institution-wide AI strategies.

The OfS's guidance calls for:

• clear rules on permitted AI use
• oversight of AI-generated teaching materials and feedback
• steps to mitigate disparities created by unequal access to paid tools and training

The last point addresses a structural inequity: students who can afford paid AI tools and training gain an advantage that the report expects institutions to actively mitigate, not merely observe.

What happens next?

The OfS plans further structured research into the scale and nature of the change. Advance HE will launch an AI-focused Change Academy for member institutions early in 2027.

For institutions, the immediate compliance question is whether a qualification still demonstrates what a student knows and can do when AI assistance may be undetectable. The OfS frames assessment security as one part of protecting degree value, alongside the same regulatory response's demands for relevant AI skills and accessible design. Universities that wait for reliable AI detection before redesigning assessment are, on this report's own evidence, waiting for a tool the sector has concluded does not exist.