The University That Caught AI Cheating With an AI That Couldn't Tell

Madeleine was in her final year of nursing at Australian Catholic University, on placement, applying for graduate jobs, when the email landed. Subject line: "Academic Integrity Concern." Body: an accusation that she'd used AI to cheat on an assignment.

Her transcript went to "results withheld." It stayed there for six months.

The accusation was eventually dropped. Six months don't come back. Madeleine told the ABC it was part of why she couldn't land a graduate position, most nursing employers require a grad year. A withheld transcript reads the same to an employer whether the reason is fraud or a false positive. They don't ask which.

This was systemic. ACU reported nearly 6,000 cases of alleged cheating in 2024. About 90 percent of them were flagged as AI-related. Deputy vice-chancellor Tania Broadley told the ABC the numbers were "substantially overstated," while acknowledging referrals for academic misconduct had spiked the year before. Around a quarter of all referrals were dismissed after investigation.

The machine in question

The tool ACU used was Turnitin's AI detector, an add-on from a company long popular with educators, bolted onto a new job it was never built for. One paramedic student described the experience with the kind of clarity you only get from having lived it: "It's AI detecting AI and almost my entire essay was lit up in blue, 84 per cent of it supposedly written by AI." An algorithm read a paramedic student's own writing and decided a robot wrote it. The student had to prove otherwise to another human being, using evidence a robot can't produce: an alibi.

Turnitin itself is candid about the limits here. On its own website, the company cautions that the AI detector "should not be used as the sole basis for adverse actions against a statement." The manufacturer is telling you, in writing, don't do the thing ACU's students say ACU did.

To be fair to due process, ACU's academic integrity officers asked for more. They demanded handwritten and typed notes, sometimes dozens of pages, plus the accused student's entire internet search history. Proof of innocence, provided by the innocent, at their own expense, on a timeline set by someone else. One paramedic student, wrongly accused, put the objection about as well as it can be put: "They're not police. They don't have search warrant to request your search history."

It got worse

Here's the part that should sting more than the false positives themselves: ACU knew. The university had been aware of problems with the Turnitin tool for over a year before it finally stopped using it, this past March. That's a year of "results withheld" emails going out from a system the institution already had reason to distrust.

Broadley, to her credit, admitted the process was messy. She told the ABC that any case where the Turnitin flag was the sole evidence was dismissed immediately once someone actually looked at it. She also admitted the investigations "were not as always as time as they should have been", an answer that reads like someone telling the truth badly, which is its own kind of honest.

What it cost

For Madeleine, six months of a frozen transcript during the exact window when nursing graduates need one to look clean. For the paramedic student, an 84 percent AI score attached to an essay they wrote themselves, and a request for their browser history to disprove it. For ACU, a quarter of nearly 6,000 referrals dismissed on review, which means the other three-quarters were either upheld or are still, presumably, somebody's open case file.

The mechanism was mundane. It took a detection tool built to flag statistical patterns in text, treated as a verdict, by people who had a year's warning that it produced false positives and kept it running anyway. The tool did exactly what it was built to do. It just wasn't built to be right often enough to hang a nursing career on, and nobody upstream priced that in until the transcripts started piling up.

The 2am version

If you're a student reading this at 2am with a paper due and a detector score you don't trust, the lesson holds whether or not you use AI. Madeleine may never have touched it. That is rather the point. The tool now sitting between your work and your degree can't distinguish your voice from a chatbot's, and it will say so with total confidence either way. Turnitin's own site tells universities not to use its score as the sole basis for punishing you. Plenty of universities act on the vendor's product without reading the vendor's fine print.

The fixable problem underneath all of this is citations. Style detection will stay messy for years, because "does this sound like AI" is a guess dressed up as a measurement. If your paper cites a source, that source should exist, say what you claim it says and still be standing when someone checks. That's a fact you can verify before you submit, while a stylistic vibe is something a detector reverse-engineers after the fact. Checking your references against the live record leaves a bad detector free to flag your prose. It stops the much older, much more embarrassing failure mode, the one where the source you cited was never real to begin with, and someone with more power than a plagiarism checker eventually goes looking for it.

Boring is the goal. Boring is never having your transcript say "results withheld" because a machine graded another machine and nobody checked the math.