Back to blog

Why does plagiarism software still miss actual cheating?

Thắng Đoàn
Thắng Đoàn

You buy detection software expecting it to catch cheating. What it actually detects is text that looks like other text, which is a different and much narrower thing.

Detection compares text, not intent

Wikipedia's plagiarism detection article defines the field as content similarity detection: locating instances of plagiarism or copyright infringement inside a document. The software does this by matching submitted text against other documents. Cheating, by contrast, is defined by Wikipedia's academic dishonesty article as any type of cheating in relation to a formal academic exercise. One is a pattern in text. The other is a human behavior. Conflating the two is where institutions get burned.

Scale makes the gap worse

Per Wikipedia, Turnitin sells automated checks of submitted documents against its database of prior submissions and website content. As of 2025, Wikipedia reports it serves over 16,000 institutions with more than 71 million enrolled students. At that scale, no human reads most work. So the similarity score quietly becomes the de facto definition of cheating, even though it was never designed to measure dishonesty. The tool's limits become your policy's limits.

Human checks do not scale either

Wikipedia's plagiarism detection article notes that human detection, the traditional method, is lengthy and time-consuming, and produces inconsistencies across an organization. So you cannot simply add reviewers on top of the software. You get either inconsistent human judgment or consistent but narrow automated matching. Neither one reliably answers whether a student actually cheated.

What this does not solve

These sources say nothing about detecting AI-generated text, false positive rates, or detector accuracy. Wikipedia's plagiarism detection entry covers text matching only. Turnitin's entry describes a similarity service founded in 1998, before generative AI existed. Any claim about AI-detection reliability would be invented here, so it stays out. The evidence does not answer that question, and neither should you trust tools that claim it without proof.

If your honor code were audited tomorrow, would it define cheating in behavioral terms or only in text-matching terms?
Which decisions at your organization are quietly delegated to a tool's output format?

In under 10 minutes, write a one-sentence definition of cheating at your organization that does not mention text matching, and check

Sources

Wikipedia: Content similarity detection

Wikipedia: Turnitin

Wikipedia: Academic dishonesty

Share:

Recommended for you

Enjoyed this article?

Subscribe for new articles. No spam. Unsubscribe anytime.

By subscribing you agree to receive the newsletter. No spam, and you can unsubscribe anytime.