What Do AI Detectors Look For? The Signals, Named and Explained

Ask someone how an AI detector works and you will usually get a shrug and a vague gesture toward "the algorithm." But detectors are not mystical. They are pattern-matchers, and the patterns they match are, for the most part, nameable. Each one is a measurable property of text that tends to look one way when a language model wrote it and a different way when a person did. The catch, and it is a large catch, is that "tends to" is doing an enormous amount of work in that sentence. Almost every signal a detector hunts for shows up in human writing too, sometimes strongly. This piece walks through the tells one at a time, names them plainly, gives you a concrete sense of what each looks like on the page, and is honest about where each one falls apart. If you understand the individual signals, you stop treating a detection score as a verdict and start treating it as what it actually is: a weighted sum of guesses.

The statistical tells: how predictable is the writing?

The oldest and most influential family of signals is statistical. These do not care what your text says. They care about how the words are arranged relative to what a language model would expect. Two terms dominate this family, and you will see them cited in nearly every detector's marketing copy, usually without much explanation. Let us fix that.

Perplexity: is every word the "expected" word?

Perplexity is a measure of surprise. Feed a passage to a language model word by word and, at each position, the model has a probability distribution over what comes next. If the actual next word is one the model rated as highly likely, the model is not surprised. If the next word is one the model rated as improbable, the model is surprised. Perplexity aggregates that surprise across the whole passage. Low perplexity means the text kept choosing the words the model expected. High perplexity means the text kept veering into choices the model would not have predicted.

Here is why this matters for detection. A language model generating text is, almost by definition, a machine that selects high-probability words. That is what it was trained to do. So model-generated prose tends to sit in a low-perplexity valley: every word is a reasonable, expected, unremarkable choice. Consider a sentence like "The results demonstrate a significant improvement in overall performance across all tested conditions." Every single word there is the safe pick. Nothing jars. A detector reading that low, smooth perplexity curve leans toward "machine."

Now the false-positive problem, and it is severe. Plenty of human writing is also low-perplexity. A competent business analyst, a policy writer, a student who has absorbed the house style of academic English, a non-native speaker who learned to write with careful, textbook-correct constructions, all of these people produce prose full of expected words in expected slots. They are not being lazy; they are being clear and conventional. Meanwhile, a novelist reaching for a surprising image, or a caffeinated blogger with a chaotic voice, will produce high-perplexity text that reads as more "human" to the detector, even if a model actually wrote it after being told to sound quirky. Perplexity measures conventionality, and conventionality is not the same as machine authorship. We go deeper on the mechanics of this in our explainer on perplexity, burstiness, and watermarks.

Burstiness: does the rhythm rise and fall?

Burstiness is perplexity's companion, and it deals with variation rather than average predictability. Human writing tends to be bursty in the sense that its sentence lengths and complexity swing around. A person might write a long, winding sentence packed with clauses, then follow it with three words. Then a medium one. Then another long one. The rhythm is uneven because human attention is uneven, because we pause, emphasize, backtrack, and change our minds mid-paragraph.

Machine text, especially from earlier or more cautious models, tends toward the opposite: a steady, even cadence where sentence after sentence lands in roughly the same length band. The prose feels metronomic. Nothing spikes, nothing collapses to a fragment. Detectors quantify this as low variance in sentence-level metrics and treat that flatness as a machine tell. High burstiness reads as human; low burstiness reads as synthetic.

And again, the exceptions swallow the rule. Technical documentation is often deliberately uniform, because uniformity aids comprehension, and a good technical writer will keep sentences in a tight, predictable band on purpose. Legal writing does the same. So does a lot of corporate communication that has been sanded down by style guides. On the flip side, modern models can be prompted to vary their rhythm, to throw in fragments, to open with a short punchy line. Once you can instruct a model to be bursty, burstiness stops being a reliable divider. It becomes a hint, weakly correlated, easily faked, and easily triggered by an innocent human who simply writes evenly.

The word-choice tells: which vocabulary shows up?

Beyond raw predictability, detectors look at the actual composition of the vocabulary. Which words appear, how often, and how they cluster. Language models have statistical habits at the level of individual word selection, and those habits leave a residue.

Over-reliance on safe, high-probability words

Because a model tends to pick likely words, it tends to lean on a particular register of vocabulary: words that are common, broadly applicable, and rarely wrong. Terms like "significant," "various," "crucial," "essential," "leverage," "delve," "robust," "comprehensive," and "furthermore" show up at elevated rates. None of these words is bad. Each is fine in isolation. But when a passage is saturated with this connective, all-purpose vocabulary and thin on specific, concrete, unexpected nouns, a detector notices the distribution skewing toward the safe center of the language.

Think of it as a fingerprint of blandness. The model reaches for the word that fits the widest range of contexts because that word carries the lowest risk of being wrong, and across thousands of small decisions the text drifts toward a beige, all-occasions vocabulary. A human expert writing about, say, sourdough will spill out concrete, domain-specific words, "levain," "autolyse," "oven spring," "the crumb," because they actually know the thing. The model, unless prompted hard, hovers at a more generic altitude.

But watch the false positive. A generalist writer covering an unfamiliar topic also reaches for safe, generic vocabulary, because they too are hedging against being wrong. A tired writer does it. A writer working in their second language does it, heavily, because the safe words are the ones they learned first and trust most. Penalizing safe vocabulary systematically disadvantages exactly the people who have the least room to be flamboyant with English. This is one of the quiet reasons detectors misfire on non-native writers, a bias we cover in our piece on why detectors produce false positives.

Vocabulary distribution and lexical diversity

A related signal is how varied the vocabulary is overall, sometimes measured as a type-token ratio or similar diversity metric. Some models, in some settings, recycle their favorite phrasings and cycle through a narrower band of words than a person might, producing text that is internally repetitive in a subtle, hard-to-spot way, the same transitions, the same sentence openers, the same handful of intensifiers. Detectors can pick up on that narrowness.

Yet the direction of this signal is unstable. Sometimes AI text is less diverse; sometimes, when a model is trying to sound impressive, it is more diverse than a plain-spoken human would ever be, stuffed with elevated synonyms nobody uses in conversation. Human writing spans an enormous diversity range too, from the deliberately spare to the ornate. A metric that points in both directions depending on the prompt is not a metric you can trust on its own.

The structural tells: how is the writing built?

Zoom out from individual words to the architecture of the piece, and another set of signals appears. These are about shape, regularity, and the scaffolding that holds a text together.

Sentence uniformity and structural regularity

Closely related to burstiness but distinct from it, structural regularity is about the internal blueprint of sentences, not just their length. Model text often falls into repeated grammatical shapes: subject, verb, object, tidy modifying clause, full stop. Paragraphs of a similar length. A rhythm of "topic sentence, three supporting sentences, transition" that repeats with almost architectural discipline. When you read enough of it, the regularity starts to feel like wallpaper, a pattern tiling across the whole page.

Detectors can approximate this by looking at the consistency of syntactic structures and paragraph proportions. The more uniform the blueprint, the more the text leans machine in the model's judgment. Human writing, by contrast, tends to have lumpy architecture, a paragraph that runs long because the writer got excited, a one-line paragraph for emphasis, a tangent that breaks the pattern.

The exception is, once more, all the human writing that is deliberately regular. Instructional content, how-to guides, structured reports, and anything written to a rigid template will show exactly this regularity, because the template imposes it. The five-paragraph essay taught in schools is a machine for producing structurally regular human text. Blaming regularity on AI ignores the vast amount of human prose that was engineered to be regular for good pedagogical or professional reasons.

The absence of human error and idiosyncrasy

This is one of the more interesting signals because it works in reverse. Detectors, formally or informally, notice what is missing. Humans make small, characteristic messes: a comma splice here, an inconsistent hyphenation there, a sentence that trails off, a mixed metaphor, a slightly wrong word that is close enough, a personal aside that has nothing to do with the argument. These idiosyncrasies are the texture of real authorship. Model text, especially unedited model text, is often unnaturally clean. It does not fumble. It does not contradict itself in that small, human way. It does not suddenly become weirdly passionate about a side point.

So a passage that is flawlessly smooth, error-free, and tonally even, never breaking character, never getting bored, never leaving a loose thread, can read as machine precisely because of its polish. The very perfection is the tell.

This one has a cruel false-positive edge. The humans whose writing is cleanest are the professionals: editors, career writers, people who proofread, people who use grammar tools. Punishing clean prose means punishing skill. A meticulous writer who has scrubbed every error out of a piece has, from the detector's point of view, made their work look more artificial. And conversely, models can be prompted to introduce faux-errors, a deliberate typo, a casual aside, so the absence-of-error signal is both unfair to careful humans and trivially defeated by anyone trying to evade detection.

Overly smooth transitions and formulaic scaffolding

Language models are exceptionally good at connective tissue. They glide from point to point with transitions that are almost too helpful: "Moreover," "In addition," "However, it is important to note that," "Building on this," "In conclusion." The seams between ideas are sanded perfectly smooth, and the overall structure often follows a recognizable arc, a tidy introduction that previews the points, a body that hits each one in order, a conclusion that restates them. It is the shape of a competent essay rendered with suspicious reliability.

Detectors and human readers alike register this formulaic scaffolding. When every paragraph opens with a signposting transition and the whole piece marches through a predictable arc without any detour or surprise, the writing feels assembled rather than thought. Real human argument tends to be messier: it doubles back, it buries the lede, it forgets to signpost, it ends abruptly because the writer ran out of steam.

Predictably, the false positive is enormous. Signposted transitions and clear structure are exactly what every writing teacher, every style guide, and every SEO best-practices article tells people to use. Web writing in particular is trained to be scannable and well-signposted. A human blogger following standard advice will produce precisely the smooth, formulaic scaffolding the detector flags. Good structure is not evidence of a machine; it is evidence of following instructions that both humans and machines have absorbed.

The surface tells: phrasings, punctuation, and hidden marks

Some signals live right on the surface of the text, in the literal characters and stock phrases. These are the ones people notice with the naked eye, and they are the ones most loaded with folklore.

Giveaway phrasings and hedging patterns

Certain constructions have become cultural shorthand for "an AI wrote this." The hedged opener: "It is important to note that." The dutiful both-sides framing: "While there are many perspectives, it ultimately depends on." The tricolon that models love: "not just X, but Y and Z." The self-aware summary: "In summary, we have explored." The relentless helpfulness that answers questions nobody asked. These phrasings emerge because models are trained to be balanced, cautious, and complete, and that training pushes them toward the same rhetorical moves again and again.

A detector, or a suspicious editor, can pattern-match on these. Cluster enough of them and the text starts to smell synthetic. But every one of these phrases was invented by humans and is used constantly by humans. Hedging is what careful people do when they are not sure. "It is important to note" is a phrase that appeared in millions of human-written documents long before any model existed, which is exactly why models learned it. Flagging the phrase flags the humans who taught it to the machine. These are among the weakest signals precisely because they are so imitable and so common in ordinary careful prose.

Punctuation and formatting artifacts

Then there is punctuation, and here the em-dash has become almost a mascot for AI suspicion. Some models produce em-dashes at a higher rate than casual writers do, along with a particular fondness for the paired-dash aside. People have started treating a dash-heavy paragraph as a confession. Curly "smart" quotes, consistent spacing, and clean formatting can feed the same suspicion.

This is thin ice. The em-dash is a legitimate, beloved punctuation mark used heavily by many excellent human writers, essayists especially. Plenty of people have always written this way. Judging authorship by punctuation preference is closer to superstition than analysis, and it produces confident, wrong accusations against people whose only crime is liking a dash.

More substantive than punctuation style are the genuine artifacts: invisible characters. Text copied out of some AI interfaces can carry non-standard whitespace, zero-width characters, or unusual Unicode that a plain human keyboard rarely produces. These are not stylistic; they are literal residue from the pipeline the text passed through, and they can be a stronger technical clue than any word choice. But they are also fragile, since a single pass through a plain-text editor strips them, and they can be introduced or removed deliberately. If you want to understand and clean these up, we have a practical walkthrough on removing hidden characters from AI text.

Watermarks, briefly

A watermark is different in kind from everything above. The signals so far are inferences drawn from text that was written normally. A watermark is a signal deliberately baked in at generation time. The idea is that a model can subtly bias its word choices according to a secret pattern, nudging toward certain tokens over equally good alternatives in a way that is invisible to a reader but detectable by anyone holding the key. If the pattern is present, you have strong evidence the text came from that specific watermarked model.

In principle this is the cleanest signal of all, because it is a designed marker rather than a statistical guess. In practice, watermarking is patchy: not all models use it, the pattern can be weakened by paraphrasing or editing, and only the party who owns the key can verify it, which is not the third-party detector most people actually use. So while watermarks matter and may matter more over time, they are not what a typical off-the-shelf detector is checking when it gives you a percentage. For the fuller picture of how they work and their limits, see our overview of how AI detection works.

The deep tell: stylometric fingerprinting

What a stylistic fingerprint captures

Stylometry is the granddaddy of authorship analysis, older than language models by decades. It rests on the observation that every writer has a measurable statistical signature, not in the fancy words they choose but in the boring ones. How often you use "the." Your ratio of "however" to "but." The average distance between commas. Your habit of starting sentences with conjunctions. The rate of function words, the distribution of sentence-initial parts of speech, the little unconscious tics you cannot help. These features are hard to fake because you are not aware you produce them, and they have been used to attribute disputed historical texts and unmask anonymous authors.

Applied to AI detection, stylometry asks whether a text's fingerprint matches the known statistical signature of a particular model's output. Models have fingerprints too, shaped by their training, and a stylometric detector can compare the two. This is often the most sophisticated layer, because it operates below the level of conscious content and can pick up on regularities the writer never intended to reveal.

The limitations are structural. Stylometry needs a decent amount of text to stabilize, so it is weak on short passages, a paragraph, a comment, an email, where there simply is not enough signal to fingerprint reliably. It also assumes fingerprints are stable, but human style shifts with mood, topic, genre, and audience, and model fingerprints shift with every new release and every prompt. A fingerprint that identified last year's model may miss this year's. And crucially, human fingerprints and model fingerprints overlap, because models learned to write from human text, so a human whose natural style sits close to the "average" learned style will register as machine-like through no fault of their own.

Why it all adds up to probability, not proof

Step back and look at the whole catalogue. Perplexity measures conventionality. Burstiness measures rhythmic variation. Vocabulary distribution measures safeness of word choice. Structural regularity measures architectural sameness. Absence of error measures polish. Smooth transitions measure formulaic scaffolding. Giveaway phrasings and punctuation measure surface habits. Stylometry measures unconscious signature. A real detector does not rely on any single one. It combines many of them, usually feeding the features into a trained classifier that has learned, from examples, roughly how to weigh them together into one number.

Notice what every signal in this catalogue has in common. Not one of them is exclusive to machines. Every single tell, without exception, appears abundantly in human writing: the conventional writer, the non-native writer, the careful professional, the template-follower, the dash-lover, the signposting blogger. Each signal is a correlation, not a certainty, and correlations stack into a probability, never a proof. That is why detectors return a percentage rather than a yes or no, and why that percentage is so easy to misread. It is a confidence estimate about a fuzzy pattern, not a measurement of a fact. We unpack exactly what those numbers do and do not mean in our guide to what detector scores actually mean.

The honest framing is this. AI detectors are looking for the statistical shadow that machine generation tends to cast: predictable, even, safe, polished, smoothly connected, structurally regular text with a certain fingerprint. When they find that shadow, they raise their estimate. But human beings cast the same shadow all the time, on purpose and by nature, and machines can be coached to stop casting it. So the signals are real, the measurement is real, and the inference from measurement to authorship is genuinely uncertain. Knowing the named signals does not just help you understand a score. It helps you see why no score, however high, should ever be treated as the final word about who, or what, wrote a piece of text.