Hunting AI cheats at university: half of those accused suffer extreme anxiety, study reveals

By Zendoric · September 8, 2026.

A survey at two Australian universities asked 771 students and identified 22 who had gone through a formal investigation for alleged academic misconduct. More than half of them —according to the study by Guy Curtis, of the University of Western Australia, and Amanda White, published in The Conversation— recorded "extremely severe" levels of anxiety: a score in the top 2% of the general population. Half also showed symptoms of severe depression. Only 3 of the 22 students stayed within the "normal" range for anxiety, depression and stress.

The researchers compared these results with those of more than 1,000 students who had not gone through misconduct proceedings and with psychological norms for more than 2,000 Australian adults collected in earlier studies. Finding accused students is difficult for academic research —disciplinary proceedings are confidential by design, to protect the fairness of the process— so the team first launched a broad survey on knowledge of plagiarism and, at the end, asked who had ever been investigated. Only in this way was it able to isolate this sample of 22 cases.

The study arrives at a time when, as the authors themselves note, many students report being accused of using generative artificial intelligence (AI) to cheat on the basis of detection tools that are not reliable. The researchers also found that a larger share of those accused did not have English as a first language, a finding consistent with earlier work on the overrepresentation of international students in misconduct cases.

One student, whose case dragged on for about six weeks, described it this way: "I felt severe anxiety throughout that whole period […] with exam stress and the case, I felt unable to function. I was already living away from home at the time and had to move back temporarily because I couldn't function due to the anxiety." Another summed up the underlying problem: "I didn't learn much because my university wasn't very kind in the process." When the procedure was perceived as punitive, the authors note, distress increased; when it was used as an educational opportunity —teaching how to cite properly, for example— some students did acknowledge learning from the experience.

The authors propose three concrete lines of improvement: redesigning assessments to reduce the opportunity to cheat with AI (more supervised exams, in-class presentations), guaranteeing access to psychologists and counselors during the investigation —with a specific effort toward international students, who are less likely to seek psychological help— and shortening the timelines of the proceedings, which are themselves an added source of distress.

This study documents with data something we had been sensing in more qualitative terms about the transition to AI in education: the friction is not only about jobs or money, it is also institutional and human, and it lands unevenly. Many universities hastily adopted AI detectors of questionable reliability to paper over a real problem —the undeclared use of generative AI in coursework— and the cost of that haste is being paid by specific students, disproportionately those who were already at a disadvantage, such as those whose first language is not English.

As sector context, we have already argued at Zendoric that the sensible response to AI in education runs through redesigning assessment around teachers who orchestrate the tool, not through policing its use with unreliable detectors turned into academic law enforcement. This study lends empirical ammunition to that thesis: chasing the symptom with an imperfect technology produces measurable harm concentrated among the most vulnerable, while redesigning the system —assessing the process and not just the final product— attacks the cause without that human cost.

In the short term, this is exactly the kind of friction to be expected in any poorly implemented technological transition: institutions deploying detection tools without the safeguards, psychological support and process review that such a deployment requires. None of this calls into question AI's potential as a personalized tutor or as an accelerator of learning, an area where we remain optimistic over the long term; what it calls into question is the haste with which it has been used as a surveillance tool before it was reliable. Institutional design, not the model, is what is failing here, and it is also the only thing universities can fix right now.