This article is written for lecturers, professors and curriculum leaders in higher education.
It's Sunday night and you're working through a pile of reflective essays. The third one is polished, well structured and cites the right theorists. So is the fourth. So is the fifth. By the eighth, you've noticed they all end with the same tidy three-part conclusion, and none of them sounds like the student who sits at the back of your Tuesday seminar and rarely says a word.
On Monday you ask that student to walk you through their main argument. They look at you, then at their laptop, then back at you. Nothing.
If that feels familiar, you're in good company. When the Associated Press spoke to US professors earlier this year, the phrase that kept coming up was "perfect homework, blank stares". Educators are tired of playing detective, and plenty are quietly asking whether a take-home written assignment can still tell them what a student knows.
This article looks at what the research says about student AI use, why detection tools haven't solved AI cheating, and what a more workable approach to assessment looks like: one built on conversation, practice and evidence you can trust.
Nearly all of them. The HEPI and Kortext Student Generative AI Survey 2026 found that 95% of full-time UK undergraduates use AI in some form, and 94% use generative AI to help with assessed work. In 2024, the figure for any use was 66%.
Most of that use isn't wholesale copying. The most common use in the HEPI survey is getting concepts explained, and 12% of students say they paste AI-generated text straight into assessed work. That 12% is up from 3% in 2024, though, and the line between "AI helped me understand this" and "AI did the thinking for me" is blurry for students and staff alike.
Faculty see the effects every week. In a survey of 1,057 US faculty by the American Association of Colleges and Universities (AAC&U) and Elon University, published in January 2026, 78% said cheating had increased on their campus since generative AI arrived, and 57% said it had increased a lot. In the UK, a Guardian investigation using Freedom of Information requests found almost 7,000 proven cases of AI-related cheating in 2023-24. That's 5.1 cases per 1,000 students, up from 1.6 the year before, and experts quoted in the investigation called those proven cases the tip of the iceberg.
Meanwhile, the people expected to respond often feel they're doing it alone. In the same AAC&U and Elon survey, 68% of faculty said their institution had not prepared them to use generative AI in their teaching. Most are redesigning assignments on their own time, between marking and everything else.
The instinctive response has been to reach for detection software. The trouble is that it isn't reliable, and when it gets things wrong, students pay the price.
In September 2025, the University of Waterloo switched off the AI detection feature in Turnitin. Its published rationale cites research on detector unreliability and bias against students writing in a second language, and notes that internal testing more than once flagged human-written text as 100% AI-generated. Waterloo concluded that no product on the market can effectively identify generative AI, and told instructors to put their energy into assessment redesign and AI literacy instead.
The bias problem is well documented. A 2023 Stanford study published in Patterns ran essays written by non-native English speakers through seven popular GPT detectors, and more than 61% of them were wrongly labelled as AI-generated. The HEPI 2026 survey also picked up rising anxiety among students about being falsely accused, including from students who say they have never used AI to write an assignment.
Detection also puts educators in the role of investigator, and it asks a less useful question. "Did a machine write this?" matters far less than "Can this student actually do what the assignment was meant to show?"
Cheating is the visible symptom. The deeper worry for most educators is what happens to learning when the thinking gets outsourced.
In a 2025 MIT Media Lab study, "Your Brain on ChatGPT", 54 participants wrote essays using ChatGPT, a search engine or no tools at all. In the first session, 83% of the ChatGPT group could not accurately quote a single sentence from the essay they had just finished. The researchers call this "cognitive debt": convenience now, paid for later with weaker recall and a weaker sense that the work is yours. It's a small preprint study, but it lines up with what faculty report. In the AAC&U and Elon survey, 90% of faculty said generative AI will diminish students' critical thinking, and 83% said it will shorten attention spans.
Add the constant pull of phones, feeds and notifications, and you get the student many lecturers describe: in the room, half there, and used to getting answers in seconds.
There is a flip side, though. The same students who drift away from a 3,000-word essay will spend hours in environments that respond to them. The Entertainment Software Association's 2025 Essential Facts report found that 60% of US adults play video games every week, and 70% of Gen Z players have met people through games they wouldn't otherwise have met. Students haven't lost the ability to concentrate. They give their attention to things that react to what they do, let them fail safely and try again, and feel like they matter.
Many faculty have pulled assessment back into the room: handwritten exams, in-class essays, cold-calling. Supervised conditions have a real place. On their own, though, they don't fix disengagement, and they add to the workload of people who are already stretched.
Tyton Partners' Time for Class 2026 report, produced with D2L and the Gates Foundation, compared faculty who redesign assessment around AI ("Integrators", 24% of faculty) with those reverting to blue books and proctored formats ("Defenders"). Integrators reported fewer attendance problems. Only 13% of faculty at institutions that ban AI entirely said their policy was effective, compared with 30% at institutions where AI is built into assignments.
The takeaway isn't "anything goes". It's that bans treat AI as a behaviour problem, when for assessment it's mostly a design problem.
The institutions making the most progress tend to share three habits.
They separate learning from verifying. The University of Sydney's "two-lane approach" puts every assessment into one of two lanes. Lane 1 tasks are secure and supervised, and verify that students can meet learning outcomes themselves. Lane 2 tasks are open, and AI use is expected and scaffolded as part of learning. Sydney's own guidance gives an interactive oral, such as a mock client conversation, as a valid Lane 1 task. Clear labels also help students, many of whom tell HEPI they aren't sure what's allowed.
They assess people, not just products. Interactive oral assessments (often shortened to IOs) were developed at Griffith University in Australia and have since been adopted by institutions including Dublin City University. They are unscripted, scenario-based conversations that mirror a real professional situation, so the assessor responds to what the student actually says. That makes them very hard to outsource. Griffith's researchers also found that the more an assessment resembles real-world practice, the less likely students are to feel exam anxiety or cheat.
They make the task worth doing. Students cut corners on work that feels pointless. Scenarios that look like the job they're heading for give them a reason to engage, and they build skills employers care about. In NACE's Job Outlook 2026 Spring Update, employers ranked communication among the most important skills for new graduates, and said it's one of the areas where new graduates fall short.
The catch is time. A 15-minute oral for each of 200 students is 50 hours before anyone has written a word of feedback. That's why so many faculty who like the idea of oral assessment never run one. It's also why some are now using AI to make conversation-based assessment practical again: the AP reported that an NYU Stern professor ran an AI-powered oral exam as his course's final assessment.
Keep the essay, but attach a five-minute conversation where the student explains one decision they made and answers a follow-up question. A Cornell biomedical engineering professor introduced an oral defence for exactly this reason, and Cornell's Center for Teaching Innovation now runs an oral assessment workshop for faculty.
Instead of "write a report on handling a customer complaint", ask students to handle the complaint. A nursing student might explain a care plan to a worried relative; a business student might pitch to a sceptical client; an apprentice might talk a new colleague through a safety procedure. You assess the knowledge and the communication skill in one go, and it's the kind of evidence employers recognise.
This is where AI becomes part of the answer rather than the problem. In Bodyswaps, students speak out loud with an AI character in a realistic scenario, such as a job interview, an anxious patient or an upset customer. You set up to 10 assessment criteria per scenario, and the AI rates each student's performance against them. Every session produces a learner report with ratings, written feedback and the full transcript, so you can see exactly what was said. Run it under supervision in class as a secure task, or set it as practice between sessions. Because the character responds to what each student says, the conversation goes wherever their answers take it. It's far harder to hand a live, spoken back-and-forth to a chatbot than a written essay, and students can practise as many times as they like before the version that counts.
Curious what that looks like from the student's side? Try Bodyswaps free and run a scenario yourself in a few minutes.
Ask students what they noticed about their own performance and what they'd do differently. In Bodyswaps, an AI coach leads a short reflection at the end of every roleplay, using questions you can set, and it appears in the learner report. Comparing how a student rated themselves with how they actually performed tells you a lot about their self-awareness, and it builds a habit they'll need at work.
If students spend their evenings on their phones and in games, put practice there too. Short, game-like scenarios that run on a phone, laptop or VR headset, with instant feedback and the chance to retry, use the same loop that keeps people playing. The difference from scrolling is that every attempt builds a skill. The Bodyswaps training library covers interviews, teamwork, difficult conversations and more, so you don't have to build everything from scratch.
A single grade tells you where a student landed. Repeated practice against the same criteria shows how far they've come. Class-level views show which criteria a whole group is struggling with, so you can adjust your teaching before the final assessment rather than after it. It's one reason universities and further education colleges use Bodyswaps to evidence soft skills development across whole cohorts.
Generative AI isn't going anywhere, and neither is the need to know what students can actually do. Detection tools can't give you answers you'd be comfortable defending at a misconduct panel. Blue books on their own don't bring disengaged students back. Conversation does better on both counts: it's hard to fake, and it builds the communication skills students will need long after they've handed in their last essay.
For years, the problem with assessing through conversation was that it didn't scale. That's no longer true. Start small: add a talk-back to one assignment, or replace one written task with a scenario this term, and see what you learn about your students.
See it for yourself: try Bodyswaps free and build a scenario for your own course, or book a demo with our team to see how assessment reports work for a whole class.
How can I tell if a student used AI to write their work?
Often you can't tell just by reading it. AI checkers make mistakes. A better test is to ask the student to talk about their work. If they can explain their ideas and answer a follow-up question, they understand it.
Do AI detectors work?
Not well enough to trust on their own. Studies show they can mark human writing as AI. They wrongly flag students who write in a second language more often. Some universities, like the University of Waterloo, have turned them off.
What is an AI-resistant assessment?
It is a task that is hard to hand over to AI. Good examples are live talks, oral exams and roleplays. In these, students have to think and answer in the moment.
Can AI help me assess my students' speaking skills?
Yes. In AI roleplay tools like Bodyswaps, students talk to a virtual character. The tool rates them against goals you set. You get a report and a full transcript for each student.
American Association of Colleges and Universities and Elon University Imagining the Digital Future Center (2026). The AI Challenge: How College Faculty Assess the Present and Future of Higher Education in the Age of AI. January 2026. https://imaginingthedigitalfuture.org/wp-content/uploads/2026/01/Elon-AACU-faculty-AI-survey-full-report-1-21-26.pdf
Associated Press (2026). Perfect homework, blank stares: Why colleges are turning to oral exams to combat AI. Published via Fortune, 25 March 2026. https://fortune.com/2026/03/25/oral-exams-colleges-anti-ai-teaching-method-gen-z-stare/
Entertainment Software Association (2025). 2025 Essential Facts About the U.S. Video Game Industry. June 2025. https://www.theesa.com/annual-esa-study-reveals-video-games-universal-appeal-across-generations/
Griffith University (2020). Less anxiety and cheating with authentic exam model. Griffith News, 21 May 2020. https://news.griffith.edu.au/2020/05/21/less-anxiety-and-cheating-with-authentic-exam-model/
Kosmyna, N. et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab, arXiv preprint 2506.08872. https://arxiv.org/abs/2506.08872
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E. and Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7). https://doi.org/10.1016/j.patter.2023.100779
National Association of Colleges and Employers (2026). Employers Say New Grads Are Largely Prepared for Work, With Room to Improve. Job Outlook 2026 Spring Update, April 2026. https://naceweb.org/career-readiness/trends-and-predictions/employers-say-new-grads-are-largely-prepared-for-work-with-room-to-improve
Sotiriadou, P. et al. (2025). Interactive orals offer a solution to AI over-reliance in higher education. Times Higher Education Campus. https://www.timeshighereducation.com/campus/interactive-orals-offer-solution-ai-overreliance-higher-education
Stephenson, R. and Armstrong, C. (2026). Student Generative AI Survey 2026 (HEPI Report 199). Higher Education Policy Institute and Kortext. https://www.hepi.ac.uk/reports/student-generative-ai-survey-2026/
The Guardian (2025). Revealed: Thousands of UK university students caught cheating using AI. 15 June 2025. https://www.theguardian.com/education/2025/jun/15/thousands-of-uk-university-students-caught-cheating-using-ai-artificial-intelligence-survey
Tyton Partners (2026). Time for Class 2026: The AI Tipping Point: From Monitoring Students to Engaging Them. June 2026. https://www.globenewswire.com/news-release/2026/06/15/3311604/0/en/Tyton-Partners-Releases-2026-Time-for-Class-Report-The-AI-Tipping-Point-From-Monitoring-Students-to-Engaging-Them.html
University of Sydney (2025). Frequently asked questions about the two-lane approach to assessment in the age of AI. Teaching@Sydney. https://educational-innovation.sydney.edu.au/teaching@sydney/?p=21264
University of Waterloo (2025). Discontinuing use of AI detection functionality in Turnitin: September 2025. https://uwaterloo.ca/associate-vice-president-academic/discontinuing-use-ai-detection-functionality-turnitin