AI detection tools produce false accusations at rates no academic-integrity process should tolerate. The faculty who are winning this moved the goalposts instead.
Matthew Kelso, Ed.D · July 2026 · 4 min read
Every provost has now sat through the same meeting. A detector flags a paper at "98% AI." The student swears they wrote it. The tool offers no evidence, because it has none — a probability score is not a witness. Somewhere in the room, someone mentions that these tools flag non-native English writers at higher rates. The case dissolves. Multiply by a semester.
Detection puts the institution in an arms race it funds both sides of, with due-process risk attached to every flag. The alternative isn't surrender. It's designing assessments where the AI question answers itself.
Assess the process, not just the artifact. Drafts with revision history, an annotated bibliography built across weeks, a ten-minute conversation about the finished paper. A student who can defend the argument owns the argument — and the conversation takes less faculty time than the misconduct hearing it replaces.
Make the AI use visible and graded. In courses where the tools are allowed, require the disclosure: what was used, for what, and what the student changed. Grade the judgment. This produces graduates who can say something true in an interview when asked how they work with AI — which is what their employers are actually asking.
Put the policy at the course level, in the syllabus. A single institution-wide rule can't fit both a poetry workshop and a data-science capstone. What scales instead: a small set of syllabus statements — prohibited, permitted with disclosure, required — that each instructor selects and owns. Students get clarity per course instead of folklore per campus.
Faculty don't need another mandate; they need worked examples from their own discipline and a semester of cover to redesign one course. The pattern that holds up: a discipline-specific workshop, one assessment redesigned per instructor, a shared repository of what worked, and a follow-up session where the misfires get discussed honestly. Chairs who fund that sequence stop reading detector reports by spring.
We run this as a faculty cohort — course-level integration without gutting pedagogy — and the assessment-redesign session is reliably where the room turns. Skeptics don't argue with a better assignment.