What AI writing detectors actually measure (and why they get it wrong)

A student gets an email saying their essay was "flagged as 78% AI." A freelancer loses a client over a detector score they were never shown. A perfectly human blog post gets rejected because it opens with "In today's digital landscape." If you write for a living — or grade, edit, or commission writing — AI-detection scores are now part of your life, and almost nobody on either side of them understands what the number actually is.

Here is the uncomfortable truth: no tool can detect AI authorship. What detection tools measure is something humbler and more useful — a set of statistical patterns that machine-generated text exhibits more often than human text. Those patterns are real and measurable. They are also things humans do all the time. Understanding the gap between "this text has AI-typical patterns" and "an AI wrote this" is the difference between using these scores sensibly and weaponizing noise.

Pattern one: uniform sentence lengths

The most robust signal is what researchers call burstiness. Human writing is bursty: we follow a forty-word sentence with a three-word one. We ramble, then snap. Statistically, that shows up as a high standard deviation in sentence length relative to the average. Language models, left to their defaults, produce eerily even sentences — fifteen to twenty-two words, again and again, paragraph after paragraph.

Measuring this is straightforward: compute the length of every sentence, take the standard deviation, and normalize it against the mean. A high ratio means human-like rhythm; a low one means uniformity. Of all the signals in this family, uniformity is the hardest for casual AI use to hide, because it emerges from how models sample text rather than from any particular word choice. But note what it can't do: a human writing in a deliberately even register — technical documentation, legal boilerplate, a careful non-native speaker — will score "uniform" too.

Pattern two: the transition tics

Every heavy user of chatbots recognizes the house style: "Furthermore," "Moreover," "It's important to note," "In today's digital age," "delve into," "a tapestry of," "plays a pivotal role," "testament to." These phrases exist in human writing — that's where the models learned them — but models reach for them at a rate no human editor would tolerate. Counting how many sentences lean on this specific phrase inventory, relative to total sentences, yields a transition rate that separates default-settings AI prose from most edited human writing.

This signal is also the easiest to defeat, which is worth being honest about: one find-and-replace pass removes it. A low transition rate proves nothing. A high one mostly proves the text was never edited by someone allergic to "moreover."

Patterns three, four, and five: hedging, repeated phrases, and a shrinking vocabulary

Three more signals round out the picture. Hedging — "might," "could," "typically," "generally," "tend to" — appears at elevated rates in model output because models are trained to avoid overcommitting. N-gram repetition counts how often the same three-word phrase recurs; models circling a topic reuse exact phrasings ("the data shows that… the data shows that") more than humans, who unconsciously vary them. And vocabulary repetition — one minus the type-token ratio — captures how quickly the text recycles its word stock. Machine text over a few hundred words tends to draw from a narrower active vocabulary than a human writing on the same subject.

Each of these is weak alone. A cautious academic hedges constantly. A recipe repeats 3-grams by necessity. A children's book has low vocabulary diversity on purpose. The signals only become interesting in combination, weighted for the kind of writing being examined — academic prose should be judged more on structural uniformity and phrase repetition, marketing copy more on transition tics and hedging, technical writing with more tolerance for repeated terminology.

What a composite score means — and what it never can

The AI text pattern detector combines exactly these five measurements — sentence-length uniformity, transition-phrase rate, hedge rate, 3-gram repetition, and vocabulary repetition — into a 0–100 score, with weighting presets for general, academic, marketing, and technical writing. Below 40, the text shows human-typical variety; 40–64 means some patterns are present; 65 and up means many of the measured patterns are firing at once. Alongside the score you get the individual metrics and plain-language flags telling you which patterns triggered, because the breakdown is more actionable than the headline number.

What the score is not — and this applies to every tool in this category, including the expensive ones — is evidence of authorship. The false-positive problem is not a bug to be patched; it is structural. Detectors are pattern-matchers trained on the overlap between two distributions that genuinely overlap: formulaic humans look like machines, and lightly-edited machines look like humans. Studies have repeatedly shown detectors flagging non-native English speakers at wildly elevated rates, precisely because careful, conservative construction is the pattern being measured. No score, from any vendor, should be treated as proof in a decision that affects a grade, a job, or a reputation.

How to actually use one

Used honestly, a pattern score has two good jobs. If you're evaluating text someone else wrote, treat the score as a prompt for questions, never a verdict — look at the flagged patterns, then look at the writing history, the drafts, the writer's other work. One signal among many.

If you're checking your own writing, the tool is more straightforwardly useful: it's a style linter. A high uniformity reading means your rhythm has flatlined — vary your sentence lengths. A high transition rate means you're leaning on connective filler — cut half of it and let the logic carry itself. Heavy hedging reads as timid whether a human or a model produced it. The patterns associated with AI writing were bad style before language models existed; models just industrialized them.

Paste at least thirty words — ideally a few hundred, since every one of these statistics stabilizes with length — into the AI text pattern detector, pick the mode that matches the genre, and read the flags before the score. If you want to go deeper on individual signals, the hedge word detector isolates the hedging axis and the sentence length analyzer does the same for burstiness. And whichever side of the score you're on, remember what it measures: patterns, not people.