top of page

Office for Students AI assessment guidance puts coursework under pressure

52 minutes ago
11 min read
Illustrative photograph. Photo: Matheus Bertelli / Pexels.

The Office for Students is pressing England’s universities to expand AI-free assessment while continuing to teach students how to use artificial intelligence. Published November 11, the Office for Students AI assessment guidance confronts a widening gap: polished coursework does not necessarily demonstrate a student’s understanding.

The central conflict concerns what a university can establish from submitted work. An essay can satisfy marking criteria while revealing little about its author’s independent reasoning. Conversely, using AI during preparation does not automatically mean a student has avoided learning or broken assessment rules.

An earlier University of Reading experiment exposed the difficulty of identifying AI submissions through ordinary marking. England’s regulator now puts the responsibility on institutions to design assessments that establish student achievement. That creates pressure on universities, assessment software suppliers, and students navigating inconsistent expectations.

Office for Students AI assessment guidance targets evidence of learning

The regulator’s intervention challenges the evidence behind a grade, while leaving room for assessments that permit AI.

The OfS published its institutional AI research with Advance HE on November 11. Conducted between June and September 2026, the project involved conversations with over 300 students and staff across more than 120 institutions.

Its findings describe concerns about unsupervised assignments that AI can complete without students demonstrating the intended knowledge. Participants also described beneficial study uses and uncertainty about assessment policies. These accounts identify institutional problems; they do not establish a national cheating rate.

The widely circulated phrase “close to impossible” reflects a staff participant’s concern about assessing students in some areas. Treating that statement as a universal finding would erase qualifications that matter. The report does not establish that every essay is AI-generated or that every coursework grade lacks value.

The accompanying assessment compliance guidance explains how existing quality conditions apply. B1 concerns academic experience, B2 addresses resources and support, and B4 covers assessment and credible awards.

The guidance supports secure assessments where necessary. These prevent AI use and include examinations, presentations, oral examinations, performances, and practical tasks. It also permits AI where the assessment still measures the intended learning outcomes.

That distinction changes the institutional question. A university cannot settle the issue simply by declaring itself supportive of AI or committed to academic integrity. It needs a defensible explanation of what each assessment establishes about the student.

Consider an illustrative economics assignment asking students to evaluate a policy proposal. A convincing final essay might show sound argumentation, or it might reflect extensive assistance. Without further evidence, the marker faces uncertainty about the student’s contribution.

Changing the font, adding a declaration, or reminding students about misconduct does not resolve that uncertainty. Those measures communicate expectations, but they do not necessarily improve the evidence supporting the grade.

An assessment that requires students to explain their assumptions creates a different basis for judgment. The instructor can examine why a model was selected, which evidence matters, and how competing explanations were rejected.

The practical significance of the Office for Students AI assessment guidance therefore extends beyond the format of an examination. Its central demand is that institutions connect the work they grade with the capabilities they claim to certify.

That connection also explains why coursework remains part of the discussion. Extended projects can reveal research, revision, and sustained reasoning. The challenge is preserving those strengths while establishing whose understanding the finished work represents.

Universities must defend what their grades actually certify

The immediate pressure falls on institutions that award qualifications, including those with clear policies but weak evidence of student achievement.

The OfS has linked assessment quality to its regulatory responsibilities. In its AI assessment announcement, it identifies potential responses to institutional failures, including conditions, monetary penalties, and funding restrictions.

Those are regulatory tools, not evidence that a university has already received a penalty over this report. The distinction matters because a warning establishes expectations without establishing an enforcement outcome.

Josh Fleming, the OfS director of strategy and delivery, called for greater use of AI-free assessment. Advance HE deputy chief executive Kathryn Harrison-Graves emphasized institutional direction, consistent guidance, and resources for staff and students.

These positions address different parts of the same problem. Controlled assessment can strengthen confidence in individual performance. Institutional coordination helps determine where that approach belongs and whether departments can deliver it consistently.

For university leaders, the resulting pressure reaches beyond academic misconduct teams. Assessment design affects staffing, teaching schedules, room availability, accessibility, and the relationship between course objectives and final awards.

An oral examination requires someone qualified to ask useful follow-up questions. A practical task requires suitable conditions and consistent marking. A revised assignment requires instructors to agree on what counts as successful reasoning.

These are implementation considerations rather than measured costs from the report. They explain why adopting a new assessment policy is easier than demonstrating that the policy works across a university.

For instructors, the shift also changes the role of marking. Reading a final answer remains necessary, but confidence increasingly depends on evidence gathered around that answer.

An instructor might compare a project with a student’s explanation of its central decision. Another might assess an applied task during a supervised session. Neither approach guarantees fairness without suitable criteria.

For students, predictability becomes essential. A course that encourages AI-supported exploration during teaching must explain the boundaries around assessed work. Otherwise, students must infer rules from scattered comments and inconsistent examples.

That uncertainty creates problems even when everyone intends to comply. Assistance with grammar, source discovery, argument development, and complete drafting involves different contributions. A single instruction to use AI responsibly leaves those differences unresolved.

Employers also have a stake, although the report does not establish how hiring practices will change. A qualification communicates a claim about a graduate’s capabilities. Assessment uncertainty weakens the usefulness of that claim.

The relevant distinction is between producing an acceptable result and possessing the judgment needed to evaluate it. Many professional tasks involve both. Universities face pressure to show which capability a particular grade represents.

For technology suppliers, the implication is similar. Features that accelerate assignment production do not automatically strengthen educational evidence. A product’s usefulness for studying and its suitability for assessment require separate judgments.

AI fluency and independent competence require different evidence

A student can learn to work effectively with AI while still needing to demonstrate understanding without it.

The apparent contradiction in the regulator’s position disappears when assessment objectives become explicit. Using a tool competently and understanding a subject independently are related capabilities, but they require different evidence.

An illustrative programming course makes that distinction concrete. A student might use AI to suggest code, then inspect its behavior and identify a security weakness. That exercise can test judgment about generated output.

A separate task might require the student to trace a simple program without assistance. Its purpose would be to establish foundational understanding. Success on the AI-assisted exercise would not automatically establish success on the independent task.

Neither task needs to represent the entire course. Together, they can answer questions that a polished final program leaves unresolved.

Earlier evidence shows why that distinction deserves attention. In a 2024 exam infiltration experiment, University of Reading researchers submitted AI-generated answers into undergraduate psychology examinations.

The university reported that 94% went undetected and that the generated submissions received higher grades on average than real student submissions. The experiment concerned a particular examination system and subject context.

It does not establish a current detection rate across universities. It does show why ordinary grading cannot automatically establish authorship or independent understanding.

Reading’s experiment and the OfS intervention approach the issue from different directions. The experiment tested whether submitted answers would pass through an existing system. The regulator asks institutions to make their assessment systems defensible.

That shift favors clearer distinctions between different types of achievement. An AI-assisted report can demonstrate evaluation, verification, and responsible tool use. An independent explanation can establish whether the student understands the reasoning behind it.

The tension becomes sharper when assessment rewards the appearance of expertise. Fluent writing, orderly references, and confident conclusions can be useful qualities. They become misleading when markers treat them as sufficient evidence of subject knowledge.

A model can produce a plausible explanation that its user cannot defend. A student can also produce awkward prose while understanding the underlying issue. Assessment needs to distinguish those possibilities without assuming either one.

The same distinction matters to knowledge workers outside education. An AI-generated summary can help someone review information, but accepting its conclusion requires checking its evidence.

For readers exploring personal knowledge management, the connection is practical: organized material supports understanding when people can examine its origins and reasoning.

That observation does not establish that any particular software improves grades or verifies authorship. It identifies a shared problem across studying and professional work: accessible information and dependable understanding are different achievements.

The strongest interpretation of the regulator’s approach therefore preserves both expectations. Graduates should understand their subjects and develop judgment about AI. An assessment system should reveal those capabilities rather than collapse them into a single finished document.

Returning to exams leaves important questions unanswered

More controlled assessment can improve confidence in individual performance without settling every question about fairness or educational value.

A supervised examination limits opportunities for outside assistance. It does not automatically establish that the task measures the most relevant capability.

Consider an illustrative research methods course. A timed test might assess whether students recognize weaknesses in a study design. A sustained project might reveal whether they can formulate a question and revise an approach.

Both forms of evidence can matter. Replacing the project entirely would change what the course measures, even if the replacement made unauthorized AI use harder.

This is the strongest skeptical challenge to a broad return to examinations. Universities need evidence that revised assessments preserve the learning outcomes they intend to certify.

Security and validity therefore require separate scrutiny. Security concerns the conditions under which a student completes the task. Validity concerns whether the task measures the intended knowledge or skill.

An institution can strengthen one while neglecting the other. A tightly controlled memory test offers limited evidence about abilities that require extended investigation. An ambitious take-home project can leave independent contribution uncertain.

The design problem involves balancing those limitations. Claiming that traditional examinations solve AI assessment would oversimplify the task. Claiming that workplace AI use makes independent assessment unnecessary would make a similarly unsupported leap.

Accessibility adds another constraint. The OfS guidance recognizes reasonable adjustments for disabled students and warns against placing disproportionate weight on AI detection outputs.

Those safeguards matter because neither a change in writing style nor a detector score explains how a student completed an assignment. An allegation and an assessment judgment require appropriate evidence.

A university should also distinguish assistance from substitution. A student who uses permitted support while developing an argument occupies a different position from someone submitting work they cannot explain.

The boundary depends partly on the assignment. Assistance acceptable during language practice might defeat the purpose of a task designed to assess unaided writing. General institutional rules need assessment-specific meaning.

There is also a risk of shifting the evidentiary burden too far onto students. Requiring extensive records of every drafting action might generate documentation without establishing understanding.

A complete activity history would still need interpretation. It might show that a student interacted with a tool, but interaction alone would not reveal whether learning occurred.

These are reasons to evaluate redesign carefully, rather than reasons to preserve every existing assignment. Institutions need to explain the contribution each component makes to a credible judgment.

The unresolved issue is therefore broader than whether examinations return. The Office for Students AI assessment guidance leaves universities responsible for showing that their chosen mixture measures achievement fairly.

Its effectiveness will depend on implementation evidence. A revised handbook can state stronger expectations; only the operation of the assessment can show whether those expectations hold.

Student adoption makes blanket assumptions unreliable

Widespread AI use establishes the scale of the assessment challenge, while leaving individual conduct and learning outcomes unresolved.

HEPI’s March 2026 student AI survey offers useful context. Conducted by Savanta in December 2025, it collected responses from 1,054 full-time UK undergraduates.

The survey found that 95% reported using AI in some way, while 94% reported using generative AI to help with assessed work. Separately, 12% reported directly including AI-generated text in assessed work.

Those categories measure different behavior. Assistance with assessed work encompasses more activity than inserting generated text. Neither figure, by itself, establishes misconduct under a student’s institutional rules.

The survey also found that 65% said assessment had changed significantly in response to AI. It therefore provides context for institutional adaptation already underway, alongside the concerns raised in the November research.

These findings undermine a simple division between AI users and independent learners. A student can use AI while engaging deeply with material. Another can use similar tools to avoid engaging with it.

The educational difference lies in the work the student still performs. Asking for an explanation can begin a productive investigation. Accepting the explanation without checking it can end that investigation prematurely.

An illustrative history assignment shows how similar actions can produce different outcomes. A student might ask AI to suggest interpretations, then compare them with primary documents.

Another student might submit the suggested interpretation without examining its evidence. The initial tool interaction looks similar, but the resulting basis for a grade differs.

That is why adoption statistics cannot settle assessment design. They indicate how widely institutions encounter AI, while offering limited information about what a particular student has learned.

The same caution applies to reports of peers completing assignments quickly. Such accounts identify a credible concern worth investigating. They do not measure how frequently the behavior occurs across the sector.

Institutions therefore need evidence from their own courses. Where do submitted answers leave understanding uncertain? Which assessment components resolve that uncertainty? Which changes preserve useful opportunities for research and revision?

Those questions also matter for AI product evaluation. A vendor might demonstrate that its software generates study questions or summarizes documents. That demonstration would not establish improved independent performance.

Educational buyers should distinguish product activity from educational achievement. Frequent use, attractive output, and student satisfaction can inform decisions, but they answer different questions from demonstrated learning.

For students, the practical implication is a clearer relationship between preparation and proof. AI can participate in preparation where permitted. Assessment must still establish the abilities the course promises to develop.

For universities, that means abandoning blanket assumptions in both directions. Widespread use does not prove widespread cheating, and an AI-friendly policy does not prove that assessment remains credible.

Watch assessment changes, regulatory evidence, and staff support

The next test is whether universities can show that redesigned assessments reveal learning more clearly than their previous arrangements.

Over the coming months, the first signal to watch is assessment-specific change. Institutions should describe which tasks require independent performance, which permit AI, and what each component establishes.

A general announcement about responsible AI would provide limited evidence of progress. Revised assignments, clear criteria, and explanations of their purpose would offer a firmer basis for judgment.

This signal would strengthen the case for effective adaptation if universities preserve meaningful coursework while addressing uncertainty about student contribution. A simple increase in examinations would leave the educational question open.

The second signal is how the regulator evaluates institutional evidence. Its published regulatory next steps say it will use existing tools rather than introduce new regulations.

The OfS also identifies expectations under a proposed future quality approach. Those expectations should remain distinct from a final implementation timetable or an announced enforcement finding.

Watch for clearer examples of what adequate evidence looks like. Institutions need to understand how assessments demonstrate achievement, rather than infer compliance from whether a particular format appears traditional.

Evidence-based scrutiny would strengthen the intervention. An emphasis on visible restrictions without examining learning would weaken the argument that the changes protect qualification quality.

The third signal is practical support for staff. Advance HE has announced an AI-focused Change Academy for member institutions in early 2027.

Its significance will depend on whether institutional planning reaches course design and teaching practice. Shared principles become useful when instructors can apply them to actual assignments and explain them to students.

Support should also connect expectations across a course. Students need to understand why assistance is encouraged in one task and restricted in another. Consistency comes from a clear educational purpose, rather than identical rules everywhere.

The Office for Students AI assessment guidance deserves attention because it connects AI policy with the credibility of grades. Its success remains an empirical question: do revised assessments provide better evidence of student achievement?

For students, instructors, and educational buyers, the next step is to examine the assessment details behind the announcements. Ask what a task establishes independently, what it establishes with assistance, and how those judgments connect.

Preserve useful distinctions when reading adoption statistics or evaluating new tools. AI assistance, generated submissions, misconduct, and weak learning describe different things. A credible response needs evidence about each.

As universities revise their approaches, read the assignment rules alongside the marking criteria. What must the student understand well enough to explain, verify, or perform, and where does the assessment actually show it?

Give every agent the context to do better work

Connect your agents to the knowledge, decisions, and history already organized in remio.

remio currently supports Windows 10+ (x64) and Macs with Apple silicon.

Your AI Partner at Work
Get more done with remio

Plan. Create. Deliver.
All in one place.

bottom of page