Why AI Detectors Flag Your Writing: Perplexity and Burstiness
Detection tools do not read your work. They measure two statistical properties of it, and both of them punish plain, careful writing no matter who typed it.
Paste a paragraph into a detector and you get back a number that looks like a verdict. Eighty-seven percent AI. It has the confidence of a lab result. It is closer to a weather forecast made by somebody standing indoors.
Knowing where that number comes from is useful. Not because you should game it, but because it explains why two tools disagree about the same paragraph, why text written entirely by hand gets flagged, and why "write more like a person" beats any advice framed around detection.
In this article
- What perplexity measures
- What burstiness measures
- Why human writing gets flagged
- Why two detectors disagree
- What a score is actually worth
- What to do instead
1. Perplexity: how surprising is the next word
Perplexity measures predictability. Feed a sentence to a language model one word at a time and ask, at each step, how likely the model considered that word. If the next word is nearly always the expected one, perplexity is low. If the text keeps taking turns the model did not see coming, perplexity is high.
Picking likely words is essentially what a language model does. So its output naturally sits at low perplexity, because it is the statistical middle of everything it has read. Detectors use that fact. Low perplexity gets treated as evidence of machine authorship.
Take the fragment "the weather was cold and". A model weights windy, wet and grey heavily. A person might write "cold and personal about it". High perplexity, and also the kind of phrase a reader remembers.
2. Burstiness: does the rhythm move
Burstiness measures variation, mostly in sentence length and complexity across a passage. Human writing is bursty. We produce a long winding sentence with three clauses hanging off it, then a short one, then something in between. The variance is high.
Model output is smooth. Sentences cluster around one length and one structure, paragraph after paragraph. Low variance, low burstiness, and a second signal for the detector.
| Signal | Typical AI text | Typical human text |
|---|---|---|
| Perplexity | Low, predictable word choice | Higher, with odd specific words |
| Burstiness | Low, even sentence lengths | High, lengths jump around |
| Vocabulary | Narrow and safe | Uneven, jargon next to slang |
| Structure | Symmetrical, heavy on triplets | Lopsided, shaped by the argument |
3. Why human writing gets flagged
Here is the uncomfortable part. Low perplexity and low burstiness are not unique to machines. They describe a great deal of perfectly good human writing:
- Technical documentation, where the whole point is that the next word is the expected one.
- Legal and academic prose, written to a rigid house style with standard sentence shapes.
- Writing by non-native English speakers, who often work from a smaller, safer set of constructions. This is a well documented source of false positives, and it has cost real students real marks.
- Anything edited hard for clarity. "Keep sentences short and use common words" is a description of low perplexity and low burstiness. Follow that advice well and a detector may decide a machine wrote it.
None of this is an edge case. It follows directly from what the tools measure. A detector cannot tell disciplined from generated. It sees statistics, not authorship.
4. Why two detectors disagree
Each tool uses a different reference model to compute perplexity, a different mix of features, a different training set and a different threshold for calling something machine-written. The same paragraph can come back at thirty percent on one and eighty-five on another. Neither number describes an objective property of your text. Each describes how that text looks against one particular set of assumptions.
Length matters too. Below roughly a hundred words the statistics are so noisy that a single sentence swings the result. Most vendors admit this somewhere in their documentation, usually well below the large percentage on the screen.
5. What a score is actually worth
A detector score is a weak signal, not evidence. Read sensibly, a high score says the text is stylistically flat and could use another editing pass. That is worth knowing whoever wrote it. Used as a verdict on authorship, especially one attached to a grade or a job, it is unreliable enough that several universities have pulled back from using it, and at least one major vendor withdrew its own detector after conceding it did not work well.
So treat the number as a readability check wearing a lab coat. It is telling you the prose has no texture. Fix the texture and the score moves, because the two things are measuring the same underlying flatness.
6. What to do instead
Write for the reader and the statistics look after themselves. The properties detectors hunt for happen to be the properties that make text pleasant to read:
- Vary sentence length on purpose. Put a four-word sentence next to a thirty-word one. That raises burstiness and improves pacing at the same time.
- Add specifics. Names, numbers, dates, places. Specifics are less predictable than abstractions, and they are the difference between a paragraph that informs and one that fills space.
- Cut the scaffolding. "It is important to note that" may be the most predictable string in AI prose, and it never carries meaning.
- Take a position. An argument somebody could disagree with is by definition not the statistical average of everything written on the subject.
- Read it aloud. The places you stumble are the places where the rhythm went flat.
If you want the mechanical half handled for you, that is what an AI humanizer is for. It restores variation in rhythm and vocabulary and strips the formulaic phrasing while keeping your meaning, which takes seconds instead of an afternoon. The specifics and the opinion are still yours to supply. There is a walkthrough in our guide on how to humanize AI text, and the tool itself is linked from the home page.
One caveat worth stating plainly
Understanding detection does not change the rules you work under. If your university, employer or publisher has a policy on AI-assisted writing, that policy applies whatever any detector says. Knowing how the scoring works helps you interpret a number somebody hands you. It is not a licence to misrepresent who wrote something, and we would not pretend otherwise.
Back to the blog · Read next: How to humanize AI text: 7 habits to edit out