Nov 2, 2026 · 6 min read
AI detectors promise a simple answer to a hard question: did a person write this? Under the hood, they don't know who wrote anything. They estimate how closely a piece of text matches the statistical patterns of machine writing. Understanding those patterns explains both how detectors work and why they misfire.
Perplexity: how surprising is each word?
Perplexity measures how predictable a text is to a language model. If every word is the one a model would have guessed, perplexity is low. Language models generate text by choosing likely words, so their output tends to have low perplexity. People make less predictable choices, so human text usually scores higher.
Burstiness: does the rhythm vary?
Burstiness looks at how much predictability changes from sentence to sentence. Human writing is bursty: a plain sentence, then a vivid one, then a fragment. Machine text tends to be uniform. Detectors such as GPTZero have described perplexity and burstiness as core signals.
Classifiers trained on examples
Other detectors, including those built into plagiarism platforms, train a classifier on large sets of known human and known AI text, then score new text by similarity. The idea is the same: they learn the fingerprint of machine writing, not the identity of the writer.
Why detectors produce false positives
Because detectors measure style rather than authorship, careful human writing can look "too predictable". Formal academic prose, technical documentation and text by non-native English speakers often uses common words and regular structure. A 2023 Stanford study found popular detectors frequently misclassified essays written by non-native English speakers as AI-generated. OpenAI withdrew its own AI text classifier in 2023, citing its low rate of accuracy.
Detectors also disagree with each other. The same paragraph can score as human on one tool and AI on another, and results shift as detectors are retrained.
What this means if you write with AI
If you use AI to draft, the detectable part is the style, not the ideas. Varying sentence length, replacing stock phrasing and adding your own perspective all move text away from the machine fingerprint, which is exactly what humanizing does. Tools that rely on tricks such as hidden characters or deliberate misspellings are easy to catch and make the writing worse.
No method can promise a particular score on every detector, and you should always follow the AI rules of your school or workplace. But writing that genuinely reads like a person is the most reliable result, both for detectors and for the people reading it.