How Do AI Detectors Actually Work? A Teacher’s Guide

If you’ve been hearing about AI detectors in the staffroom and wondering what’s really going on under the hood, you’re not alone. I spent a few weeks digging into this and wanted to share what I found.

At their core, most AI detectors work by analyzing patterns in text. They look at something called “perplexity” and “burstiness.” Perplexity measures how predictable the word choices are. AI-generated text tends to pick the most statistically likely next word in a sequence, which makes it very low perplexity. Human writing is messier, more surprising, and tends to score higher.

Burstiness is about variation in sentence structure. Humans naturally write with a mix of short punchy sentences and longer, more complex ones. AI tends to produce more uniform sentence lengths and structures.

Tools like Turnitin, GPTZero, and Originality.ai each use slightly different models, but they’re all trying to answer the same question: does this text look like it was generated by a large language model?

Here’s the thing though. These tools aren’t perfect. They produce false positives (flagging human-written text as AI) and false negatives (missing AI-generated text that’s been lightly edited). The accuracy varies a lot depending on the type of writing, the student’s proficiency level, and whether the text has been paraphrased.

I think every teacher using AI detection should understand these basics before relying on the results. What’s your experience been? Have you had any surprising results from AI detectors in your classroom?