Advisory: AI Writing Detection at Mount Royal University

Generative AI in Teaching and Learning Working Group
August, 2026

Summary

The Genenerative AI in Teaching and Learning Working Group recommends that Mount Royal University avoid the use of applications to detect AI-generated content including, but not limited to, Turnitin's AI writing detection feature for use in academic integrity processes. Our recommendation draws on peer-reviewed studies, the official statements of comparable universities, and the vendor’s own documentation. We organise the reasoning into four areas: the tool’s empirical reliability in both directions of error, and more broadly equity and inclusion concerns, fit with our evidentiary standards for misconduct, and effects on classroom culture and the instructor-student relationship.

1. The tool gets both kinds of errors wrong

A detection tool can fail in two directions. It can flag work by students who did not use AI (a false positive) and it can pass work by students who did use AI (a false negative). Independent peer-reviewed evaluations document both problems at rates that make the tool unsuitable for academic integrity decisions.

On false positives, Weber-Wulff and colleagues (2023) tested fourteen detectors, Turnitin included, in the International Journal for Educational Integrity and found that none reached eighty percent overall accuracy. Walters (2023), comparing sixteen detectors in Open Information Science, recorded false-positive rates on human-written control essays ranging from zero to thirty-one percent across the tools (although Turnitin fared better than most in their test). McGill University’s official guidance reports that “OpenAI states that their false positive rate is 9%, which is a similar rate found in other AI detection tools.” OpenAI itself withdrew its own AI Text Classifier in July 2023, citing low accuracy.

On false negatives, Perkins and colleagues (2024), in the International Journal of Educational Technology in Higher Education, tested seven detectors against AI text that had been lightly modified using simple techniques students would intuitively try: varying sentence length, adding spelling errors, paraphrasing with a free tool. Baseline accuracy across the tools was 39.5 percent. After these modifications, average accuracy dropped by 17.4 percentage points. Turnitin specifically fell from 50 percent accuracy on unmodified AI text to 7.9 percent on modified text, a 42.1 percentage point drop and the largest of any tool tested. Sadasivan and colleagues (2023, updated 2025 in Transactions on Machine Learning Research) provide theoretical work showing that detection accuracy approaches chance as language models improve, an effect they also demonstrate empirically.

The practical consequence is that students willing to make minor edits will likely pass through the tool undetected, while students who write carefully and formally without using AI are exposed to false flags. This pattern of errors is the opposite of what a tool meant to support academic fairness should produce.

 

2. Equity and inclusion: predictable harms to identifiable groups

A well-cited study on detector bias by Liang and colleagues (2023), published in Patterns (Cell Press) tested seven widely used AI detectors on essays by non-native English writers (drawn from the TOEFL pool) and by US eighth-graders. Across the seven tools, an average of 61.3 percent of the non-native essays were flagged as AI-generated, while the native-speaker essays were flagged at near-zero rates. Eighty-nine of the ninety-one non-native essays were flagged by at least one detector, and eighteen were flagged unanimously by all seven. The reason is technical rather than behavioural: detectors examine how predictable a text is, and second-language writers tend to use a narrower vocabulary and more regular sentence structures, which the tools read as machine-generated. When the research team rewrote the same essays to use a broader vocabulary, misclassification dropped by nearly 50%, confirming that the bias is a property of the writing style the tools penalise, not of the writer's intent.

The same mechanism predicts disparate impact on other identifiable groups, including students writing in technical or scientific registers where convention rewards predictable structure, neurodivergent students whose academic prose tends toward formal regularity, and students who use assistive writing tools. Turnitin acknowledges that content produced by Grammarly’s generative features will likely be flagged by its detector, though most students cannot easily tell where Grammarly’s grammar-checking features end and its generative features begin. For an institution with stated commitments to equity, diversity, and inclusion, adopting a tool whose error rates fall predictably on these groups is difficult to justify.

 

3. Evidentiary fit: the score cannot support punitive action

Mount Royal University, like most Canadian universities, decides academic misconduct on the balance of probabilities. An AI detector score cannot meet that standard for three reasons that compound.

First, the vendor itself disclaims this use. Turnitin’s official documentation states: “Our AI writing detection model may not always be accurate (it may misidentify human-written, AI-generated, and AI-paraphrased text), so it should not be used as the sole basis for adverse actions against a student.” Since July 2024, Turnitin has gone further and stopped reporting a numerical AI score at all when the result falls between zero and twenty percent, replacing the number with an asterisk in the report, because “there is a higher incidence of false positives when the percentage is between 0 and 19.” When a vendor disclaims the use case for which a tool would be most operationally valuable, and suppresses the lower portion of its own output to avoid misinterpretation, an institutional position relying on that tool becomes difficult to defend on appeal.

Second, an AI score is unverifiable. Unlike plagiarism similarity, which points to a specific document an instructor and student can both read, an AI score points to nothing checkable. UBC’s official rationale for not enabling the feature lists this as one of five concerns: “Results from the feature are not available to review.”

Third, the meaning of any individual score depends on a quantity the institution cannot measure. Bassett and colleagues (2026), writing in the Journal of Higher Education Policy and Management, work through the statistics with a representative case. Even with a generous one percent false-positive rate and ninety percent true-positive rate, the probability that an individual flagged paper is actually AI-generated ranges from below fifty percent to above ninety-five percent depending on the unknown proportion of students using AI in the cohort. The same score can mean “more likely than not” or “less than a coin flip,” and the institution has no way to tell which.

Peer institutions report that these problems matter operationally. Washington State University, which has a no-sole-evidence policy and has cancelled their Turnitin AI detection as of February 2026, has published that between 2023 and 2025, thirty-three percent of academic integrity review board cases involving alleged AI use were dismissed because AI detection was the only supporting evidence. The detector flag at best opens an inquiry; on its own it does not sustain a finding.

4. Classroom culture and the instructor-student relationship

Even when a detector result is used only to begin a conversation, the act of running detection at scale changes how instructors look at student work and how students experience submitting it. Turnitin itself has acknowledged the effect. In a company blog post on understanding false positives, Turnitin writes that “if you don’t acknowledge that a false positive may occur, it will lead to a far more defensive and confrontational interaction that could ultimately damage relationships with students.” The vendor is describing here what the academic integrity literature has documented: detection encourages a posture of suspicion, and students who know their work is being assessed by such a tool respond defensively.

There is also a subtler effect on instructor judgement. Bassett and colleagues (2026) observe that the surface features instructors are encouraged to treat as AI hallmarks (formulaic prose, predictable structures, lists of points, formal vocabulary) appear in AI text precisely because they occur in the human writing on which AI models were trained. An instructor who has a flagged score in hand is predictably more likely to interpret those features as evidence of misconduct rather than as the conventional academic prose they are. McGill’s official guidance makes the same point in plainer terms: “False positive results misguide instructors and can create situations where students are wrongly accused of a violation that they did not commit, forcing them to defend work that is rightfully theirs.” Sarah Eaton (2023) at the University of Calgary has written about how false accusations of academic misconduct are associated with reputation damage, anxiety, depression, social isolation, and other documented harms to student wellbeing.

Several major Canadian research universities have declined this tool on similar grounds. UBC chose not to enable Turnitin's AI detection in April 2023 and reaffirmed that decision in August 2023; the University of Waterloo discontinued it as of September 2025, citing internal testing in which the product flagged human-written text as 100 percent AI-generated; Western University stopped using it in January 2024 over concerns about accuracy; McGill’s official guidance discourages its use; and Dalhousie states that Nova Scotia privacy law and the university's Protection of Personal Information Policy bar submitting student work to any AI detection tool. Outside Canada, Vanderbilt disabled the feature in August 2023, UC Berkeley opted out after a pilot, Washington State University cancelled its Turnitin AI detection contract in February 2026, and Australian National University disabled the feature on January 1, 2024.

Recommendation

We recommend that instructors avoid the use of AI detection at this time, including Turnitin's AI writing detection feature. Our academic integrity response should focus on three things that the literature and our peer institutions converge on: assessment design that surfaces process and authorship (in-class writing, scaffolded drafts, oral defences); transparent statements in course outlines about what AI use is and is not permitted in each course, as our guidelines support and advocate; and academic integrity processes that emphasise instructor judgement and conversation with students rather than detector scores. A detection tool that produces both false positives and false negatives at the rates documented, that disadvantages identifiable student groups, that cannot meet the evidentiary standard our procedures already require, and that the vendor itself acknowledges can damage classroom relationships is not the tool to anchor that work.

The Generative AI in Teaching and Learning Working Group is proceeding with a recommendation to disable the AI detection tool in Turnitin, but in the interim wanted to advise faculty members of the limitations of the tool to support pedagogical decision making.

Questions? Please contact Michelle Yeo (myeo@mtroyal.ca), Chair of the Generative AI in Teaching and Learning Working Group