Academic integrity
Universities dropped AI detectors. What should replace them?
Detection asked the wrong question. The useful question is not “who wrote this?” but “what does this show about the student’s reasoning?”
Why universities are switching detectors off
Vanderbilt University disabled Turnitin’s AI detector in August 2023, noting that even a 1% false-positive rate would have wrongly flagged around 750 of the 75,000 papers it submitted the year before. The University of Waterloo discontinued the same feature in September 2025, citing unreliability, bias against students whose first language is not English, and internal tests in which human-written text was flagged as 100% AI-generated.
The harm is not spread evenly. Liang et al. (2023) found that popular detectors labelled 61.22% of essays by non-native English speakers as AI-generated. Plain, careful prose—exactly what many multilingual students write—looks “machine-like” to a statistical detector.
The gap detection leaves behind
Waterloo’s guidance is to focus on evidence of learning rather than hunting for misuse. That is the right instinct, but it leaves markers with a practical problem: large cohorts, fluent text that can hide weak thinking, and pressure to stay consistent across many hands. Marking “by feel” brings back an older bias—polish and native fluency quietly raise grades.
An alternative: mark the reasoning, with evidence
- Separate expression from reasoning. Score claim structure, use of evidence and calibration apart from grammar and style.
- Link every judgement to the text. Each mark should point to the passages that justify it, so students and moderators can see and challenge it.
- Check fairness by language and cohort. Compare outcomes across groups every cycle instead of assuming a rubric is neutral.
- Keep the final call human. Use tools for a consistent first pass and to flag uncertain cases—never for an automatic verdict.
This is the approach CERENO Eval is built around. It collects the passages relevant to each criterion, tests whether they really support it, and gives the marker an evidence-linked profile to accept, adjust or reject. It is not an AI detector and does not judge authorship.
Questions to ask before adopting any tool
- Can it show the evidence behind every score?
- Has it been compared against expert markers on your own students’ work?
- Does it report results by language background and cohort?
- Who makes the final decision, and how can a student appeal?
- Is student work ever used to train its models?