Integrity and Detection

This is the highest-stakes section on the site, because it is the only one where getting it wrong costs a child something they cannot undo.

Three things shape everything here.

For comparison with workplace systems that formalise time or activity data, Monitask’s page on stealth monitoring software shows how the same type of measurement is handled outside education.

Detector accuracy claims do not hold up evenly across students. The core finding, published in Patterns in 2023, tested seven commercial detectors against 91 TOEFL essays written under supervised exam conditions by non-native English speakers. The mean false-positive rate was 61.3%, while essays by US eighth-graders were classified almost perfectly. The tools were measuring how elaborate the prose was, not who wrote it.

A rate measured on a population is being applied to a person. When Turnitin launched its detector claiming a 1% false positive rate, Vanderbilt University did the arithmetic against its own 75,000 annual submissions — around 750 students — and disabled the tool. Dozens of institutions have since done the same, and published their reasoning.

For an external perspective on academic integrity, technology, and student rights, see International Center for Academic Integrity.

The durable answer is design, not detection. If a task can be completed acceptably by a machine that never met your students and does not know what you taught, the task was measuring something you did not intend to measure. That was true before any of this existed.

Nothing in this section is legal advice, and rules differ everywhere. What it offers is a procedure: what to check before you raise a concern, how to open the conversation, and what to say to a parent who asks whether a machine accused their child.

One principle sits underneath all of it. A false accusation of cheating is not a small administrative error — it goes on a record, it changes how a student is seen by other staff, and it teaches a young person that being honest did not protect them. Following the procedures here means you will occasionally let something through. That is the correct trade, because the other error is the one you cannot undo.

• Accusing a Student: What You Owe Them

  Before you raise it, and while you raise it. A procedure that protects the student, and also protects you if you turn out to be wrong.

• Assessment That Needs No Detection

  The detection problem is mostly an assessment design problem. Ways to set work where the question of who wrote it stops mattering.

• What Detector Accuracy Claims Mean

  A detector advertising 98% accuracy can still be wrong about your class in ways that matter. How to read the number before you rely on it.

• False Positives, and Who They Hit

  The errors are not spread evenly. They concentrate on particular students, and dozens of institutions have acted on that.

• A Classroom AI Policy on One Page

  Most students say nobody ever told them the rules. A one-page policy your class can actually follow, and how to write it yourself.

• When Work Changes Suddenly

  A jump in quality is the commonest trigger for suspicion and a poor one. The other explanations, and how to find out which applies.

• Talking to Parents About It

  Four conversations that come up, and what to say when a parent asks whether you accused their child on the word of a machine.