Everything About AI
NextFin News - Britain's higher education regulator is facing pressure to force universities back to paper-based and in-person examinations, as a new report argues that handwritten tests under supervision are now the most reliable defence against artificial-intelligence cheating. The Office for Students is being urged to require an immediate end to unsupervised remote exams, a finding that lands as generative AI use in assessed work has become near-universal: a 2026 survey of 1,054 UK undergraduates found 94% had used the technology to help with assessed work, and 12% had put AI-generated text straight into it.
The shift is not nostalgia for blue books. It is the admission of a harder truth: when a chatbot can produce a credit-worthy essay in seconds and no detector can prove it did, the only way to certify that a student did the work is to watch them do it. That realisation is redrawing the boundary between what universities assess at home and what they assess in an exam hall — and it is creating a new market for the infrastructure of secure assessment.
The Numbers Behind the Warning
The regulatory pressure follows a report from the thinktank Policy Exchange, Evidence of Learning Through Assessment: Protecting the Value of a Degree in the Age of AI, authored by Swansea University neuroscientist Professor Philip Newton and published on August 17, 2026. Freedom of Information requests sent to 120 UK universities in 2024 found that 78% were using online, remote exams for summative assessment, 70% planned to continue with the format, and only 10% used online invigilation for all such tests. The report also found that 67% of university policies governing those exams did not mention generative AI at all.
The exposure is broad. More than 90% of students report being assessed by take-home essays, and 55% by remote online exams. Adoption of the tools has raced ahead of policy: the Higher Education Policy Institute's 2026 survey found 94% of undergraduates had used generative AI to help with assessed work. An earlier HEPI survey, in February 2025, put overall student AI use at 92%, up from 66% in 2024, with 88% using it for assessments compared with 53% a year earlier. The proportion of students directly including AI-generated text in assessed work rose to 12% in 2026, up from 3% in 2024.
Disciplinary data shows the problem surfacing in cases. More than 2,000 undergraduates at Russell Group universities were punished for misusing generative AI in the 2024-25 academic year, compared with around 700 in 2023-24, according to data released under the Freedom of Information Act; the University of Glasgow, the University of Leeds, University College London and Imperial College London all expelled students. At the school-qualification level, Ofqual recorded 100 proven AI-related plagiarism cases in the summer 2025 exam series — 75% of all plagiarism cases, though only 2.0% of the 5,025 total student malpractice cases. Mobile phones, not chatbots, were involved in 44.3% of student cases.
"Too many university assessment systems can currently be completely bypassed using AI, eroding the value of a degree and undermining the role of universities," said Sir Vince Cable, the former business secretary, in a foreword to the report. "In-person examination should replace remote assessment unless the latter is fully secure."
Neil Mendoza, provost of Oriel College, Oxford, wrote in a foreword to the report: "Universities must act now to put their house in order, before it is too late." The report recommends that the Office for Students and the Quality Assurance Agency intervene to require the immediate cessation of online unsupervised remote exams, a return to in-person assessment, more detailed degree classifications showing what a graduate actually learned, and independent training standards for educators.
Why Detection Failed and Paper Did Not
The case for paper rests on a mechanism, not a preference. Assessment only works if the work submitted under a student's name reflects that student's ability. Generative AI severs that link for anything done without supervision. In a blind test cited in sector analysis, answers written entirely by GPT-4 were entered into five undergraduate psychology exams taken at home: 94% went undetected, and on average they scored half a grade boundary higher than the real students' work.
The obvious alternative — AI-detection software — has largely failed the credibility test. Several major US universities, including Vanderbilt and Northwestern, disabled Turnitin's AI-detection tool in 2023 over false-positive fears; Vanderbilt estimated the tool's 1% false-positive rate would have wrongly flagged about 750 of the 75,000 papers it processed in a year. UK regulators have drawn the same conclusion: Ofqual, the Joint Council for Qualifications, Jisc and the Office of the Independent Adjudicator all treat detector scores as at most one source of evidence, never proof. In one test of 14 detection tools, none reached 80% accuracy; in another, mean accuracy fell from 39.5% to 22.1% after simple machine edits.
That leaves two defensible options: supervise the moment of production, or make the process visible. Paper exams under invigilation do the first. They are not new technology; they are the absence of a vulnerability. This is why the regulator's intervention is being framed as a return to basics rather than an innovation — the cheapest way to guarantee authorship is a room, a desk and a person watching.
Ofqual, which regulates school qualifications in England, has been explicit about where the risk sits. Updating its AI approach on July 16, 2026, it stated that "supervised exams remain largely protected, as students are not permitted to access AI tools during exams," while non-exam assessment — coursework completed outside controlled conditions — faces greater pressure. It has also told exam boards that AI may not be used as a sole marker, and in March 2026 it wrote to boards demanding immediate steps against AI misuse, including more rigorous authentication of candidate and teacher work.
But the regulator is also constrained by evidence. Proven AI cases remain a small share of total malpractice — 2.0% in summer 2025 — and the Russell Group has said the rise in detected cases may reflect better reporting rather than a surge in misconduct. The tension is between a small number of detected cases and a large number of undetectable ones: the 100 proven AI cases tell you what was caught, not what happened.
The Regulator's Bind
The Office for Students has so far taken a principles-based approach. In a statement published on June 5, 2025, Josh Fleming, the regulator's Director of Strategy and Delivery, wrote that "resisting AI integration... would be a temporary solution with significant opportunity costs for both institutions and students," and encouraged universities to experiment with how they support students. That stance sits awkwardly with the Policy Exchange recommendation that the regulator should enforce a ban on unsupervised remote exams "if needed."
The bind is structural. The Office for Students' core regulatory condition requires each assessment to be valid and reliable, and its guidance says assessments that let students gain marks for work that is not their own would likely be of concern. Yet it has no appetite to prescribe assessment format, and universities guard their autonomy over how they teach and test. A mandate to end remote unsupervised exams would be the regulator stepping into the exam hall — precisely the kind of prescriptive intervention it has avoided.
Ofqual faces a parallel question on the other side of the screen. It has not yet decided how to regulate on-screen exams; its consultation closed on March 5, 2026, and it has proposed a default of no on-screen assessment with up to two new on-screen specifications per exam board as exceptions, none in subjects with more than 100,000 entries. The Joint Council for Qualifications' exam instructions, in force from September 1, 2026, require AI to be switched off in three places — the application, the operating system and the network — and require exam centres to test their setup before the exam. The message is consistent: the device is no longer a trusted environment.
The Second-Order Effect: The Degree Signal Decouples
The first-order story is cheating. The second-order story is what happens to the credential when employers stop trusting it. If a degree no longer certifies unassisted ability, employers will build their own verification — and some already have. The report points to the Solicitors Qualifying Exam, created in 2021 by the Solicitors Regulation Authority to replace law-school-administered exams, as the template: when a sector cannot guarantee a standard, an independent gate appears outside it.
That shift has commercial consequences. A degree whose signal is decoupling from skill pushes verification spending downstream — to hiring tests, to professional licensing, and to the vendors of secure-assessment infrastructure. It also creates a two-lane assessment market. The University of Bath announced that from the 2026-27 academic year it would move to a "Two-Lane" approach developed with the Association of Pacific Rim Universities: "open" assessments where AI use is permitted and declared, and "closed" assessments where it is not. That design — secure lanes for credentialing, open lanes for learning — is likely to become the sector's operating model.
The market is already pricing the shift. The global AI-in-education market is projected to grow from $7.52 billion in 2025 to $10.6 billion in 2026, a compound annual growth rate of 40.9%, according to market research. The broader education-technology market is projected to grow from $213.2 billion in 2026 to $437.5 billion by 2033. The beneficiaries are not the detection tools that universities are abandoning, but the infrastructure of secure delivery: lockdown browsers, on-screen exam platforms with AI-switched-off configurations, and process-evidence systems that capture drafts and checkpoints.
For universities, the economics cut the other way. Supervised exams are expensive: they consume exam-hall estate, require invigilation staff, and compress marking into narrow windows. Several Russell Group universities, including the University of Manchester and Imperial College London, reintroduced in-person written exams for courses that had moved online during the pandemic, but doing so at scale means reversing years of digital-assessment investment. That is why the report pairs the call for in-person assessment with a demand for educator training in "modern, AI-proof" assessment design — the sector is being asked to spend on both the old model and the new one.
The Counter-Thesis: Why Going Backwards May Cost More Than It Saves
The strongest argument against a wholesale return to paper is not that it will not reduce cheating — it will — but that it trades one problem for several others. The Quality Assurance Agency warns that reverting to handwritten examinations would be regressive, undermining progress on accessibility. Students who rely on assistive technology funded through the Disabled Students' Allowance can be stripped of those tools by blanket no-AI rules. Handwritten exams also sample a narrower range of competencies and place heavy demands on university estate at a time when institutions are under financial pressure.
The alternative camp argues for redesign rather than retreat. Some institutions are moving to assess "process" over "output": civil-engineering students might use AI to help design a building, then verify every AI output by hand calculation; literature students might submit drafts alongside final essays so the evolution of the work is visible. The University of Lincoln has moved toward assessing students' process over their outputs from September 2026. Europe makes greater use of the interactive oral viva, which is far harder to outsource to a chatbot — though a 2026 study at New York University found that while 70% of students agreed AI-conducted oral exams tested genuine understanding, 83% found the format more stressful than a written exam, and 83% had never taken an oral exam before.
The Office of the Independent Adjudicator, which handles student complaints about universities' internal procedures, reported a "relatively low" number of academic-misconduct complaints around perceived AI use — though it acknowledged that could rise in coming years. That is the quiet counter-signal to the crisis narrative: the formal complaint machinery has not yet been overwhelmed.
The honest assessment is that both forces are at work. The return to paper is a cyclical correction — a pendulum swing back toward supervision after a decade of digital flexibility. But the underlying driver is structural: generative AI has permanently broken the assumption that an unsupervised written submission represents unaided work. That assumption will not repair itself, which means some form of authenticated, supervised, or process-evidenced assessment is here to stay. The question is not whether the unproctored essay survives; it is which combination of invigilation, oral defence and process evidence replaces it.
What to Watch
The near-term test is enforcement. Watch whether the Office for Students and the Quality Assurance Agency adopt the Policy Exchange recommendation to require an end to unsupervised remote exams, and whether Ofqual publishes its response to the on-screen consultation later in 2026. The Office for Students and Advance HE also expect to publish research on AI in higher education later in 2026.
The key falsifying signal is the data. Ofqual's malpractice statistics for summer 2026, due around December 10, 2026, will show whether AI cases are rising or plateauing. If proven AI cases stay below roughly 3% of all malpractice cases while adoption of process-based and two-lane assessment rises, the case for a blanket return to paper weakens considerably — the panic would be outrunning the evidence. If instead AI cases climb while detector reliability stays poor, the regulator's hand hardens and the paper-exam revival accelerates.
Three scenarios frame the next year. In the base case, universities adopt a two-lane model: secure, supervised assessment for high-stakes credentialing, and open AI-permitted work for learning and low-stakes tasks. In the upside case for traditionalists, the regulator mandates an end to unsupervised remote exams, triggering a wave of spending on exam halls, invigilation services and secure on-screen platforms. In the downside case, institutions overcorrect, accessibility suffers, and the sector fractures between well-resourced universities that can afford supervised assessment and the rest.
The deeper read is that this is not really about paper. Paper is just the cheapest way to solve an authorship problem that detection software could not. The lasting change is the end of the unproctored essay as a trusted signal — and with it, the beginning of a market where proof of who did the work is a product you buy, not a presumption you make.
Explore more exclusive insights at nextfin.ai.
