90% of Student Work Is AI-Made: How University Exams Will Change
Surveys show up to 92% of students use AI in their work. Here is how university exams are changing, from handwritten blue books to oral vivas.
Nine in ten students now use generative AI in their academic work. According to the Higher Education Policy Institute (2025), 92% of UK undergraduates used AI tools in their studies, up from 66% a year earlier, and 88% had used them in assessments. Faced with numbers this size, universities in the US, UK and Australia are quietly redesigning university exams rather than trying to police the technology out of existence.
The shift matters well beyond campus. Assessment is the mechanism that makes a degree mean something to employers, regulators and graduate schools. If most submitted work now involves AI, the question is no longer whether students use these tools, but which formats can still certify what a graduate actually knows.
- HEPI’s 2025 survey found that 92% of UK undergraduates used AI in their studies, up from 66% in 2024.
- Sales of handwritten exam «blue books» rose as much as 80% at large US universities between 2023 and 2025, the Wall Street Journal reported.
- Turnitin found that 11% of more than 200 million papers it reviewed contained at least 20% AI-written text (2024).
- Regulators and universities are converging on a «two-lane» model that separates secured, supervised exams from open, AI-permitted assessments.
How Much Student Work Is Really AI-Made?
According to the Higher Education Policy Institute (2025), 88% of UK undergraduates used generative AI in assessed work, up from 53% in 2024. Turnitin’s analysis of over 200 million submissions (2024) found 11% contained at least 20% AI-written text, and 3% were mostly machine-generated.
The headline «90%» figure circulating in faculty meetings blends two different behaviours. Most students use AI to explain concepts, summarise readings or structure drafts. A smaller but significant share submits substantially AI-written text as their own. HEPI found only a minority admitted pasting unedited AI output into assessed work.
The speed of adoption, rather than misconduct itself, is what has unsettled institutions.
«It is almost unheard of to see changes in behaviour as significant as this in just 12 months.»
Notably, research from Stanford Graduate School of Education (2023) found that self-reported cheating rates among US high school students stayed flat after ChatGPT launched, at roughly the same 60–70% range recorded for years. AI did not invent academic dishonesty. It changed its economics, making high-quality shortcut work effectively free and instant.
The Return of Handwritten University Exams
US universities are reviving supervised, handwritten testing at scale. The Wall Street Journal (2025) reported that sales of exam «blue books» rose about 80% at UC Berkeley, 50% at the University of Florida and 30% at Texas A&M over two years, as instructors moved graded work back into invigilated rooms.
The pattern repeats across markets. Several Australian universities expanded in-person examination after the national regulator TEQSA asked every institution for an AI risk plan in 2024. UK departments have restored closed-book finals in courses that had moved fully to coursework during the pandemic.
Handwriting is not nostalgia. A proctored room is currently the only assessment environment where an examiner can be confident no model was involved. But it measures a narrow slice of ability under artificial conditions, and it scales poorly for large cohorts and distance learners.
That tension explains why few institutions are betting everything on paper. Most are pairing secured exams with assessments that assume AI will be used, and telling students explicitly how to use AI in assignments without risking academic penalties.
The Two-Lane Model: Secured and Open Assessment
Assessment researchers in Australia formalised the emerging consensus in 2023 as a «two-lane» system. Lane one covers secured, supervised tasks that certify individual competence. Lane two covers open tasks where AI use is permitted and declared, reflecting how graduates will actually work. TEQSA endorsed the framing in its 2024 sector guidance.
Under this model, a degree programme no longer treats every essay as evidence of unaided ability. Instead, it concentrates verification in a few high-stakes moments and lets the rest of the curriculum absorb AI openly.
| Assessment format | How it verifies learning | Main trade-off |
|---|---|---|
| Proctored handwritten exam | Physical supervision removes AI access entirely | Narrow skills tested; heavy logistics for large cohorts |
| Oral exam (viva) | Live questioning probes real-time understanding | Examiner time; risk of inconsistent grading |
| Process portfolio | Drafts, version history and reflection show the work’s evolution | Process evidence can itself be AI-assisted |
| AI-integrated open task | Grades judgment, verification and editing of AI output | Requires new rubrics and disclosure norms |
Oral examination is the format gaining ground fastest in pilot programmes. A ten-minute structured conversation exposes whether a student understands the essay they submitted, regardless of who or what drafted it. Its constraint is cost: examiner hours do not scale the way automated grading does, though some institutions are testing AI-assisted viva scheduling and transcription to reduce the load.
Why AI Detection Cannot Carry the Load
Automated detection has proven too unreliable to anchor academic integrity. A Stanford study published in Patterns (2023) found that leading AI-text detectors flagged more than half of essays written by non-native English speakers as machine-generated, while OpenAI withdrew its own AI classifier in July 2023 citing low accuracy.
The asymmetry is structural. Detectors look for statistical fingerprints in text, and light paraphrasing removes most of them. False positives, meanwhile, carry severe consequences: students accused of misconduct on the basis of a probability score, with no way to prove a negative.
Several universities, including Vanderbilt in 2023, disabled AI-detection features in their plagiarism software for exactly this reason. The practical consequence is that integrity has to be designed into the assessment itself, not bolted on afterwards. Institutions now publish disclosure frameworks so students know where legitimate AI assistance ends and a violation begins, rather than leaving the boundary to a detector’s opinion.
What Changing University Exams Means for Students
The redesign of university exams shifts what gets rewarded. According to OpenAI (2025), more than a third of US adults aged 18–24 use ChatGPT, with roughly a quarter of their messages related to learning. Assessment is following usage: fluency with AI is becoming an examined skill, while unaided reasoning gets tested in fewer, more controlled moments.
For students, three practical changes follow. First, high-stakes grades will increasingly depend on live performance: closed-book exams, vivas and in-class writing. Second, open assignments will require documented process and explicit AI disclosure, with penalties attached to concealment rather than to use. Third, the study-tools market is adapting around this split, as consumer EdTech apps, from Quizlet to Spanish startups such as Modo Cheto, position themselves for AI-permitted preparation rather than answer generation.
For universities, the cost curve is the harder problem. Supervised and oral formats are expensive, and the institutions that solve verification at scale will hold a credibility advantage in a market where employers already discount inflated coursework grades.
The open question is no longer whether AI belongs in student work, but whether university exams can be rebuilt fast enough to keep degrees credible while it does. The 2026–27 academic year, the first in which most curricula were designed after mass AI adoption rather than before it, will be the first real test.
