How do AI text detectors actually work?

An AI text detector doesn't find anything in your writing. It re-reads your text with a language model of its own, scores how much it resembles writing the detector was trained to call machine-made, and prints that as a percentage. Nothing is extracted, because nothing was left behind for it to find.

Perplexity and burstiness

This is the explanation you'll see everywhere. Perplexity is predictability: for each word, how likely was it given the words before it? Models tend to pick likely words, so their output often scores low. Burstiness is how much that predictability varies across a passage — people write a long sentence, then a fragment, then a tangent, and early model output was flatter than that.

Both are real measurements, and neither is out of a person's reach. Careful, plain, heavily edited prose has low perplexity too, which is why clear writing gets flagged.

It's also out of date. GPTZero dropped perplexity and burstiness in 2023 in favour of a trained classifier, and most commercial detectors now work that way instead.

What a classifier is comparing you to

A classifier is shown a large pile of confirmed-human writing — usually text from before 2022, since that can't contain modern model output — and a large pile of generated text, and it learns whatever separates the two piles: phrasings, punctuation habits, transitions, rhythm. Given your paragraph, it reports which pile the paragraph looks more like.

That makes the score a similarity judgement about somebody's training data, not a record of how your text was produced. The vendors say as much in their own documentation: a "20% AI" result is not 20% of the document, and it is not the probability that anyone used a model.

What is checkable instead

Things that are actually in the file. Hidden characters are one — you can see exactly which codepoints are there, which is a fact rather than a score, though it still says nothing about who wrote the text. The statistical watermarks providers embed would tell you more, and you can't read them: Google's SynthID marks Gemini's text output, but its public verification portal doesn't accept text at all. Here's what exists today.