The basic idea
AI detectors try to answer a simple question: was this text more likely written by a person or generated by a language model? They cannot see who typed the words, so they rely on statistics. Language models produce text by repeatedly predicting the most probable next word. That process leaves fingerprints, and detectors are trained to spot them.
Most commercial detectors, including Turnitin, GPTZero, Originality.ai, and Copyleaks, combine two approaches. The first measures how predictable the text is. The second uses a classifier, a separate model trained on large collections of human and AI writing, that learns subtler patterns. The final score is a probability, not a verdict.
Perplexity: how surprising is each word?
Perplexity is a measure of how surprised a language model is by a piece of text. If every word is exactly what the model would have predicted, perplexity is low. Human writing tends to include unexpected word choices, personal references, and idiosyncratic phrasing, which raises perplexity.
AI-generated text usually has low perplexity because it was produced by choosing likely words. That is why phrases like "plays a crucial role", "in today's fast-paced world", and "it is important to note" are such strong signals. They are statistically safe choices that models love and that careful human writers tend to avoid.
Burstiness: how varied are the sentences?
Burstiness describes variation across a document. People write in bursts: a long, winding sentence followed by a short one. A fragment. Then a question. Language models, especially on default settings, produce sentences of remarkably even length and structure. Detectors measure that variation and treat uniformity as a sign of automation.
Burstiness also applies to ideas. Human writers drift, add asides, and return to earlier points. AI text tends to march through a topic in a neat sequence of evenly weighted paragraphs, often ending with a tidy summary that begins with "In conclusion".
Classifiers: learning the patterns
Beyond those two signals, modern detectors use classifiers trained on millions of labeled examples. These models learn combinations of features that are hard to describe in words: the frequency of certain function words, the distribution of punctuation, the way paragraphs open and close, and how often hedging phrases appear.
Classifiers are powerful, but they have a weakness. They are only as good as the data they were trained on. When a new language model is released, or when text is edited by a person, the patterns shift and accuracy drops until the detector is retrained.
Why false positives happen
False positives, human writing flagged as AI, are the biggest controversy around detectors. They happen most often with writing that is naturally formulaic: lab reports, legal summaries, short answers, and text by non-native English speakers who rely on common, safe phrasing. Several studies have found that non-native writers are flagged at much higher rates than native speakers.
Short texts are also risky. With only a few sentences, there is not enough data to measure burstiness reliably, so scores swing wildly. That is why most detectors recommend a minimum length and why many institutions say a detector score should never be the only evidence in an academic integrity case.
What this means for writers
If you use AI to help you draft, the most reliable way to produce writing that reads as human is to make it genuinely more human. Vary sentence length. Replace stock transitions with ones that follow your actual reasoning. Add specifics: a real number, an example from your own experience, a source you read. A humanizer like Humanizely automates much of the structural work, and its built-in score bar lets you check results before you share them.
If you are a teacher or editor, treat detector scores as a prompt for conversation rather than proof. Ask about the writing process, look at drafts, and consider the student's previous work. Detectors are a useful signal, but they are statistical tools with real limitations, and understanding how they work is the best way to use them fairly.