AI Detectors Are Becoming a Campus Flashpoint

1 The professor who tried everything

The Atlantic piece, published 21 September 2026 by Will Oremus, opens with Timothy Paustian, a University of Wisconsin at Madison biology professor who found early detectors comically bad, then embedded prompts only chatbots would answer.

Between hidden prompts and detectors, he caught 60 AI-written essays on one assignment last year among 350 students in his online microbiology class; all but one admitted using AI.

2 Detection is now a bought product

The magazine reports Pangram has emerged as the gold standard for detection, while higher education most commonly uses Turnitin's, an add-on to its plagiarism checker. About 1,400 North American colleges and universities have purchased it, chief product officer Annie Chechitelli told the magazine, saying it flagged at least some AI writing in nearly half of scanned work last year.

That is a company statement, and the rate depends on how a hit is counted.

3 The flashpoint: disputes and appeals

Many faculty are skeptical of detectors but at a loss for what else to do. Marc Watkins, a University of Mississippi lecturer who studies AI in education, told the magazine most faculty feel overwhelmed.

Guidance points the other way. An MIT working group issued an influential report recommending against relying on detectors, warning of an atmosphere of distrust between instructors and students. Paustian's university discourages relying on detectors alone.

4 Why detection alone fails the case

A flag is not a finding: admissions and stripped prompts are judgment calls by people who did not see the writing happen.

The magazine reports that four years into the ChatGPT era, colleges may be further from solving AI cheating than when it launched. Paustian plans to scrap writing assignments this year, telling the magazine it breaks his heart, because writing is one of the best ways to learn.

5 What assessment redesign can do

The alternative is changing what the work asks for, so a detector is not the load-bearing evidence. In-class writing, staged drafts, oral defense and tasks tied to a class's own material produce evidence about a student, not a text's statistical profile.

That route costs time and class size, which is why dropping assignments is a data point. Designing assessment around what can be witnessed and defended is more work than a scan, but it addresses what a scan cannot.

6 The bottom line

Reported: detectors are procured at scale while guidance warns against relying on them. Reported limit: flag rates come from the vendor, with no independent accuracy measure. The fix under argument is assessment design, not a better threshold.