What is a qualification supposed to prove?
When ChatGPT arrived in late 2022, universities reacted the way most large institutions react to anything new and frightening. They reached for a tool. Within months, Turnitin had bolted an AI writing detector onto the software that already sat inside most submission systems, and for a while it felt like the problem might have a tidy technical fix. Paste in an essay, get a percentage, done. It was reassuring in its simplicity, a familiar dashboard that promised to separate the authentic from the synthetic with the same cold certainty as a plagiarism check. Administrators breathed easier; assessment integrity, it seemed, had been safeguarded with a few lines of code.
It has not worked out that way.
The trouble with AI detectors is that they are guessing, and they often guess badly. They look for patterns that supposedly mark out machine-written text, things like unusually even sentence rhythm or a low level of unpredictability in word choice. The problem is that plenty of human writing looks exactly like that. A careful, methodical student who writes in plain, well-structured English can trip the same wires as a chatbot. So can someone writing in their second or third language, which is why several studies have found these tools flag work by non-native English speakers far more often than they should. One widely cited analysis from Stanford researchers found that detection algorithms misclassified over half of essays written by non-native speakers as AI-generated, a statistical bias that ought to have disqualified the tools from serious use from the start. A false accusation of cheating is not a small thing. It can follow a student through a degree and into a career, undermining their academic record and professional reputation long after the misunderstanding could have been cleared up. And detectors hand those accusations out with a confidence they have not earned, presenting fuzzy probabilities as firm verdicts and leaving students to prove a negative—that they wrote their own work—which is nearly impossible to do with any finality.
This is why a lot of the early enthusiasm has quietly drained away. Some universities have switched the AI detection feature off altogether rather than rely on numbers they cannot defend in a misconduct hearing, and even those that keep it active tend to treat its output as little more than a conversation-starter for academic advisors rather than as actionable evidence. Meanwhile, the technology itself has not stood still. Each new generation of large language model produces text that is more natural, more variable, and harder to distinguish from human writing, which means the detection arms race is one that educators were always destined to lose. Even OpenAI, the company that built the thing everyone was panicking about, pulled its own AI text classifier within a year because it simply was not accurate enough to be useful. The company’s researchers quietly admitted that the classifier caught less than 30 percent of genuine AI writing while flagging human text at an unacceptable rate. When the people who make the AI cannot reliably spot the AI, a clean detection score starts to look like wishful thinking dressed up as quality assurance.
There is a deeper issue underneath all of this, and it is the one worth talking about.
Detection treats AI as a contamination problem, something to be screened out so that assessment can carry on as before. But the way we assess has already changed, whether universities like it or not. A take-home essay written over two weeks was always a slightly artificial measure of what someone knew, more a test of time management and library navigation than of genuine intellectual mastery. Now that a student can generate a passable draft in thirty seconds, it measures almost nothing on its own. The very premise of the unsupervised, untimed, text-based assignment has been hollowed out, and no amount of software patching will put the substance back in. What we are witnessing is not a temporary disruption but a fundamental mismatch between the format of traditional assessment and the capabilities of modern tools. The question is not whether that mismatch exists, but how honestly we are willing to name it.
The universities handling this well have stopped asking how to catch AI use and started asking what they are actually trying to assess. That shift looks different on every campus. Take Queen Mary University of London, one of the larger institutions in the UK capital and the kind of place where this debate plays out in real seminar rooms rather than in policy documents. Its students are spread across East London, many of them living close enough to walk in. Sitting right on the QMUL campus next to Regent’s Canal, Scape Mile End gives East London students easy access to Shoreditch, Brick Lane and the city centre via the Central line. For a cohort that mixes home and international students from dozens of countries, blanket detection tools are a particularly blunt instrument, and the more thoughtful response has been to redesign the work itself.
What does that redesign look like in practice?
More assessment that happens in the room, where a tutor can watch the thinking happen. Oral examinations and vivas, where you have to defend an argument out loud rather than submit it and disappear. Assignments built around a specific lecture, a local data set, or a piece of fieldwork that a general-purpose model has never seen. Marking the process, the rough notes and the redrafts, rather than only the polished result. Some departments now ask students to submit annotated bibliographies and planning documents alongside final essays, not as bureaucratic hoops but as genuine windows into intellectual development. Others have reintroduced closed-book, in-person written exams for core knowledge while reserving open-ended projects for higher-order synthesis. None of this is about banning the technology. It is about setting work where using a chatbot to do the whole thing simply does not help you very much, because the value lies not in the final page but in the messy, iterative labour that produced it. When the assignment is rooted in the specific, the local, and the immediate, a general-purpose language model has little to offer beyond generic fluff.
That is a harder road than buying a detector and trusting the percentage, and it asks more of academics who are already stretched thin. Redesigning curricula, retraining staff, and rethinking assessment frameworks take time, money, and institutional will—all of which are in chronically short supply. It is much easier to click a button and get a number, even a misleading one. But it is the honest one. The students sitting in those seminars are going to spend their entire working lives alongside these tools, collaborating with them, managing them, and learning to distinguish their strengths from their profound limitations. Pretending we can wall them off for the duration of a degree was never realistic, and a detection score that might be wrong was never going to hold that wall up. The sooner we stop treating AI as an intruder to be expelled and start treating it as a fact of the intellectual environment to be navigated, the sooner our assessment practices can regain their credibility. The institutions that come out of this period in good shape will not be the ones with the best detection software. They will be the ones who are willing to ask an uncomfortable question about what a qualification is supposed to prove, and then change their teaching to match the answer. AI did not create that question. It just made it impossible to keep ignoring.
