UCT is right: universities should ditch flawed AI-detection tools

Opaque, unreliable and putting students in constant fear, such software harms education more than it protects it. By Sioux McKenna and Neil Kramm

- Sioux McKenna and Neil Kramm

The recently announced University of Cape Town (UCT) decision to disable Turnitin's artificial intelligence (AI) detection feature is to be welcomed - and other universities would do well to follow suit. This move signals a growing recognition that AI detection software does more harm than good.

The problems with Turnitin's AI detector extend far beyond technical glitches. The software's notorious tendency towards false positives has created an atmosphere in which students live in constant fear of being wrongly accused of academic dishonesty.

Unlike their American counterparts, South African students rarely pursue legal action against universities, but this should not be mistaken for an acceptance of unfair treatment.

As Rebecca Davis has pointed out in Daily Maverick: detection tools fail. The fundamental issue lies in how these detection systems operate. Turnitin's AI detector doesn't identify digital fingerprints that definitively prove AI use. Instead, it searches for stylistic patterns associated with AI-generated text.

The software might flag work as likely to be AI-generated simply because the student used em dashes or terms such as "delve into" or "crucial" - a writing preference that has nothing to do with artificial intelligence.

This approach has led to deeply troubling situations. Students report receiving accusatory emails from professors suggesting significant portions of their original work were AI-generated.

One student described receiving such an email indicating that Turnitin had flagged 30% of her text as likely to be AI-generated, followed by demands for proof of originality: multiple drafts, version history from Google Docs, or reports from other AI detection services like GPTZero.