Should Universities Just Turn the Detectors Off?

In August 2023, Vanderbilt University announced it was disabling the AI detection feature in Turnitin, the plagiarism platform it already licensed. The explanation was careful and specific: concerns about accuracy, about how the tool worked, and about false accusations landing on students.

It was one of the first prominent schools to say it publicly. It hasn’t been the last. I think more should follow, and I want to make that case without pretending AI-assisted cheating isn’t real.

The wrong kind of mistake

Every test makes mistakes. The question is which mistakes you can live with.

When a detector misses AI-generated work, a student gets a grade they didn’t earn. That’s bad. When a detector flags honest work, a student gets accused of misconduct — sometimes facing failing grades, disciplinary records or lost scholarships, on the strength of a number they can’t inspect or appeal in any meaningful way.

Those errors are not symmetric. An education system should be far more afraid of the second than the first. Detectors, deployed at scale and read by busy people, make the second kind of error routinely.

The errors aren’t random

If false positives were spread evenly, you might argue they’re an unfortunate cost of doing business. They aren’t. Research has found detectors disproportionately flag writing by non-native English speakers, and there are good reasons to expect similar effects for students taught to write formulaically or who lean on grammar tools.

A tool that systematically suspects the same groups isn’t neutral infrastructure. It’s a policy decision, made by default.

A tool that systematically suspects the same groups isn’t neutral infrastructure. It’s a policy decision, made by default.

“But it’s just one signal”

Defenders say detector scores are only one input, and instructors use judgment. In the best cases, that’s true. But a percentage on a screen has gravity. It anchors the conversation. A student trying to prove a negative — I didn’t use AI — starts from behind.

And the instructors being asked to “use judgment” are rarely trained on what detector scores mean, how they’re calibrated, or how they fail.

What to do instead

Turning detectors off isn’t giving up. It’s moving effort somewhere it works:

• Assess process, not just product. Drafts, outlines, annotated sources and in-class writing build a record of learning that a chatbot can’t fake in one prompt.
• Talk to students. A five-minute conversation about an essay reveals more than any score.
• Write clear AI policies per assignment. Students can’t follow rules nobody has stated.
• Redesign the assignments most vulnerable to automation, rather than policing them harder.

None of this is free. It takes time, and teaching time is scarce. But the alternative — outsourcing suspicion to a black box that gets the hardest cases wrong — has costs too. They just fall on students instead of institutions.