AI detectors do not "know" who wrote a text. They estimate how likely a text is to have been produced by a language model by measuring statistical properties of the writing. Understanding those measures makes the score far more useful — and keeps you from over-trusting it.
Perplexity: how predictable the words are
Language models choose words that are statistically likely. A text made of likely words is "low perplexity". Human writers make less predictable choices — an unusual verb, a specific detail — which raises perplexity. Detectors compute how surprised a model is by each word.
Burstiness: how much the rhythm varies
People write in bursts: a long sentence, then a short one, then a question. Model outputs tend to keep an even length and structure. Measuring the variation of sentence length and complexity is the second classic signal.
Trained classifiers
Modern detectors go beyond two numbers: a classifier is trained on large sets of human and generated texts and learns many subtle patterns. It outputs a probability for the whole text and, in the detector of this suite, a score per sentence.
Reading the score
A probability of 85 % does not mean 85 % of the text is generated: it means the pattern is strongly consistent with generated text. Read the sentence-level highlights; a few very mechanical paragraphs in a human text are common, and vice versa.
The limits
Short texts, lists, legal or technical prose and heavily edited drafts are often misclassified. Non-native writers can be flagged more often because their phrasing is more regular. No detector should be the sole basis for a decision about a person.
Checklist
- Analyse at least a few hundred words
- Read the sentence highlights, not just the number
- Treat the result as one signal among others
- Compare with the author's other writing when possible
- For your own drafts, rewrite the flagged passages
Used as a reading aid, a detector tells you where a text feels mechanical and why. Used as a verdict, it will be wrong often enough to matter.