Can We Still Trust AI Detectors?

The question 0:01

If software flags a student's essay as written by AI, what does that flag actually tell you about the student?

This is WAC Signal, from the World Assessment Council. Evidence and ideas for the changing world of learning.

Episode seven. Can we still trust AI detectors?

The signal 0:29

Universities are switching detectors off. Curtin University in Australia turned off Turnitin's AI detection from the first of January 2026, while keeping its text-matching. A July 2026 commentary from the Higher Education Policy Institute notes that UCLA and UC San Diego made similar decisions in 2024. In Singapore, Nanyang Technological University has confirmed it will stop using automated AI detection tools from 2027.

At the same time, the problem the detectors were bought for is growing. A Guardian investigation using freedom-of-information requests to UK universities found that proven cases of AI misconduct rose from one point six per thousand students in 2022-23 to five point one in 2023-24, almost seven thousand cases, with early figures pointing to about seven point five per thousand in 2024-25. Over the same period, proven traditional plagiarism fell.

And one detector company reports high rates of AI writing. Turnitin reported in July 2026 that nineteen point four percent of US higher education submissions to its AI detector, between October 2025 and April 2026, had an AI writing score above eighty percent, against nine point eight percent in the UK and ten point two percent in Australia. Those figures come from the same class of tool whose reliability is in question, and from a company that sells it.

So institutions face a real problem, and a tool that many of them no longer trust to solve it. The question is why.

The evidence 2:29

Start with fairness. Weixin Liang, James Zou and colleagues at Stanford ran seven widely used detectors on ninety-one TOEFL essays written by non-native English speakers. Eighty-nine of the ninety-one were flagged by at least one detector, while essays by US eighth graders were classified almost perfectly.

The cause is structural. Detectors reward varied, unpredictable phrasing, which second-language writers use less. So the students most likely to be wrongly flagged are the students least able to defend themselves.

That pattern is now showing up in formal complaints. The Office of the Independent Adjudicator for Higher Education published four case summaries on AI misconduct in July 2025. Three involved international or second-language students, and three were upheld or partly upheld against the university. International students make up twenty-four percent of all UK higher education students, and fifty-one percent of postgraduates, according to that Higher Education Policy Institute commentary.

Then there is scale. When Vanderbilt University disabled Turnitin's AI detector in August 2023, it noted that the vendor's stated false-positive rate of one percent, applied to the seventy-five thousand papers it had submitted in 2022, could have meant around seven hundred and fifty student papers wrongly labeled as partly AI-written.

And a low error rate does not tell you what a flag means. As our white paper on assessment puts it, a small false-positive rate cannot say how likely any one flagged script is to be machine-written, because that depends on how many scripts in the pile were.

What that looks like in practice was reported in Australia. In October 2025 the Australian Broadcasting Corporation reported that Australian Catholic University had registered nearly six thousand cases of alleged academic misconduct in 2024, about ninety percent of them related to AI, according to internal documents.

The university's deputy vice-chancellor, Tania Broadley, called those figures substantially overstated. She said about half of confirmed breaches involved AI, around a quarter of all referrals were dismissed after investigation, and any case resting solely on Turnitin's detector was dismissed immediately. One nursing student waited six months to be cleared. The university stopped using the tool in March 2025, after finding it ineffective.

Students feel this. In RAND's 2025 surveys, half of US students worried about being falsely accused of cheating. And in the Higher Education Policy Institute's 2026 survey of UK students, being accused of cheating remained among the most cited reasons for avoiding AI.

The tension 5:59

Detection is improving. In a September 2025 working paper, not peer reviewed, Brian Jabarian and Alex Imas tested three commercial detectors and one open-source model on everyday writing. One commercial tool, Pangram, had false-positive and false-negative rates near zero. But the benchmark was not student coursework.

And people do no better than the software. In a blind test at the University of Reading, markers missed ninety-four percent of AI-written exam answers. Experts quoted by the Guardian called the recorded misconduct cases the tip of the iceberg.

So the honest position is uncomfortable. Undetected AI misconduct is real, and switching off a detector does not make it go away. But a detector score is not reliable enough, or fair enough, to decide a student's future.

The implication 7:04

A detector can be, at most, a prompt for a conversation with a student. It is not evidence on its own. Our assessment white paper puts a simple test to any program team: if a misconduct case rested on a detector score alone, would your process dismiss it?

It also means the real fix is not in the integrity office. If a task can be completed by AI without anyone noticing, a better detector will not repair it. It is the task that no longer measures what it claims to.

The WAC view 7:43

In our AI in Learning Report 2026, we recommend that institutions stop using AI detector scores as evidence of misconduct. The error rates are high, and they fall hardest on second-language students. We also recommend a secure-plus-open, or two-lane, assessment model, with the invigilation it needs funded.

The University of Sydney's two-lane model is the best-known version. Lane one assessments are supervised, and assure that students have mastered required knowledge and skills. Lane two assessments are open, encourage responsible AI use, and should make up most of the assessment load. The university states plainly that an unsupervised assessment that forbids AI is an invalid combination, because AI use cannot be reliably prevented or detected without supervision.

Others are following. The University of Bath will move from a traffic-light AI policy to the two-lane model from 2026-27, Cardiff University said it is looking at the same model, and the London School of Economics reported more in-class, invigilated and oral assessment.

And whatever replaces detection should be checked for fairness too. Our assessment white paper asks institutions to check outcomes for second-language and disabled students in every secure format, and to count how many misconduct findings rest on a detector score alone.

What to watch 9:23

One. How many more universities switch off automated detection, and what they put in its place.

Two. Whether independent researchers test the newer detectors on real student coursework, not only on everyday writing.

Three. Complaints and appeals. Whether cases like those published by the Office of the Independent Adjudicator keep finding that second-language students carry the cost of detection errors.

The sources for every figure in this episode are listed on the episode page, along with the full AI in Learning Report 2026, at worldassessmentcouncil.org.

WAC Signal is produced by the World Assessment Council, with AI-assisted research and AI voices, and every claim is checked against its source before publication.

This has been WAC Signal. Evidence and ideas for the changing world of learning.