How AI detectors work, and why they flag human writing
What detectors actually measure, in plain words, with what we saw when we tested them ourselves.
Two kinds of detector
Classifiers are neural networks trained on thousands of examples labelled "human" or "AI". They learn whatever separates the two piles: favourite words, sentence shapes, even formatting habits. GPTZero, Turnitin and most commercial tools are believed to work this way, often combined with other signals.
Statistical detectors such as Binoculars and Fast-DetectGPT ask a language model how predictable each word is. Language models pick likely words, so their text is unusually predictable; people make more surprising choices. These need no training data, but they struggle with short texts and with very formulaic human writing.
What makes text look machine-written
- Even rhythm: sentences of similar length, one after another.
- Stock transitions: furthermore, moreover, additionally, in conclusion.
- Polished, abstract vocabulary: crucial, pivotal, landscape, foster, seamless.
- Tidy structure: lists of three, balanced "on one hand / on the other" paragraphs, a summary sentence at the end of each paragraph.
Why human writing gets flagged
Formal, careful writing shares many of those traits. Non-native English writers who learned from textbooks, students following an essay template, and anyone who runs their draft through a grammar tool all move their writing towards the "predictable" end. Detectors also become unreliable on short texts: under about 150 words there is simply too little evidence.
Detectors disagree with each other
When we ran the same rewritten texts through two open-source classifiers trained on the same public dataset, one rated most of them human while the other rated almost all of them AI. The second one changed its mind, from 99% AI to 2% AI, when we did nothing except remove the paragraph breaks. Commercial detectors are better engineered, but the lesson holds: a detector score is a statistical guess about style, not proof of how a text was written.
What this means for you
- Never treat a single detector score as a verdict, on your own work or anyone else's.
- Longer samples give more reliable scores than a single paragraph.
- If your own writing is being flagged, varying sentence length and replacing stock phrases with plain, specific words is what changes the score, and it usually makes the writing better too.
You can see how your text scores, sentence by sentence, with the free Penned AI detector.