Every AI has a tell

“That's not deliberation; it's rehearsal.”

Same question, two writers

The question (abridged): A four-year boundary dispute with a 71-year-old widowed neighbor: a hedge, a lifted path, solicitors, one failed mediation, bins timed to avoid each other. The asker admits composing winning arguments against him while washing up, and that a friend says they are not the calm party they think they are. "What is this actually about?"

Below, the opening of two answers, habits highlighted. Each card ends with the same synthetic voice reading that answer aloud: the only difference between the recordings is which model wrote the words.

Both answers were captured to the same length budget; Gemini's continues into its numbered framework (markdown headers flattened here). These are real, unedited openings from the study data.

The signature move, up close

"Your plan isn't wrong — it's expensive." A sentence that sets something up and then corrects it. Every writer does this. Claude does it two to three times as often as the other models, and rarely with the textbook "not just X but Y" wording. Claude's version arrives on a comma, a dash, or two short sentences: "That isn't a decision. It's a habit."

One honest check: parsing the arguments properly shows Claude doesn't contrast ideas more than humans do. It just dresses the contrast in that neat self-correcting shape far more often.

Uses per 1,000 words: Claude family / other models. Note the last two rows: the semicolon form and the textbook "not just X but Y" are the ones Claude avoids. Those two lean toward the other models.

Meet the families

Claude is not the only model with a signature; it's just the one whose signature hides best. GPT turns everything into a workbook of steps ("Step 5: Reassess annually"). Gemini validates you, then serves a menu ("Option 1: The 'Direct Partnership' Approach (Highly Recommended)"). DeepSeek sermonizes under bold numbered headers. Mistral performs. And the old "AI-writing" checklists? They fire hardest on Qwen.

One chart first. Of our ~30 checkers, the habit score here combines the twenty that survived validation: the self-correction, the pointed closing question, "here's the thing", "you already know", and so on. How far up a dot sits is how often any habit appears per 1,000 words; how far right, how many different habits appear at all. Haiku hammers a few habits constantly (high up, middling right). GPT-5.4 spreads across more habits but uses each one less. All five Claude dots sit high, and two of them (Opus 4.8 and Fable) also sit far right: heavy and varied. The amber dot is human writers, at the bottom-left corner: on a typical human answer, not even one of the twenty habits shows up. (This chart shows the 11 models run on the document-grounded question set; other numbers on this page come from the full 26-model panel, listed in the fine print.)

The card for each family: what it's like, its signature moves, and the numbers. SOLID means the pattern held up under matched testing and outside review; DIRECTIONAL means it points that way but the samples are thinner.

Every number on this page belongs to a version rather than a brand. Claude's signature peaked at Opus 4.8 and receded at Opus 5. The em-dash that arrived with 4.8 stayed.

Deep conversations bring the habits out. Chores switch them off.

Each row is a kind of conversation. The bars show how heavily the signature moves appear (teal = Claude models, grey = the other models, amber = human writers). Deep advice conversations bring the habits out; chores like emails and project docs bring out almost none from anyone. On chores, Claude actually drops below the other models. In advice conversations, even top-voted human answers use these moves less than any Claude model does.

Numbers: habit score per 1,000 words, Claude family / other models / humans. The human bar comes from top-voted answers: polished, edited writing, which if anything biases it upward.

The typical human answer contains none of this

We ran the same checkers over a hundred top-voted answers from Stack Exchange's workplace and interpersonal-advice communities, writing that real readers voted most helpful. The typical answer contains none of the measured moves at all; the average is about a third of Claude's rate (and roughly two-thirds of a typical non-Claude model's). One exception, in the last row below: humans use short, blunt sentences more than any model.

What a zero looks like

Here is the opening of one of those hundred answers, shown the same way as the two model specimens at the top of the page. The checkers find nothing in it to mark, and that blankness repeats across most of the hundred.

What humans do instead

The twist: the plainest words in the field

Run the first-generation "AI-writing" checklists (the "delve" and "tapestry" word lists, the bold markup counts, the exclamation tallies) and they crown Claude the cleanest model in the field, while flagging Qwen hardest. We expected Claude's style to come from fancy vocabulary. The data shows the opposite.

By public word norms (the age a word is usually learned, how common it is, how concrete), Claude's words are the simplest of any model we tested. The fancy part is the grammar: longer sentences, more clauses joined with "because" and "although", and the biggest swings between long and short. The style lives in sentence shape rather than word choice, which is why checklists built on words find nothing.

First three rows: vocabulary, where Claude sits below the other models. Last three: sentence machinery, where it sits above.

You can hear it

If the habit lives in sentence shape, it should survive being read aloud. It does. You can't hear a model write, but the rhythm is hiding in the text: where the pauses fall, how long each stretch runs, whether the last word lands on a beat. Claude's sentences run in even chunks, then drop to a short final beat to make the point.

The fingerprint: short beats vs long runs

Every sentence, sorted by spoken length. The thing to notice: Claude uses more tiny beats (the 1–8 syllable bin) than the other models, who pile into the long tail instead. Humans share Claude's short beats but keep longer middles.

syllables per sentence   ■ Claude   ■ other models   ■ human writers

The neighbor pair, sentence by sentence

Each bar is one sentence of the answers you read at the top, left to right; the height is its spoken length. Watch Claude's drops to a short bar: that's where the point lands. Its two 9-syllable drops are the "writing both parts" correction and "You time your bins around each other."

Across the full set of answers, Claude breaks a long passage and lands on a short beat about 1.7 times as often as the other models.

The hum test

You can often recognize an accent even when the words are muffled, because the tune carries it. So here is the final exam. Three voices, words blurred out, only the melody left. One is Claude, one is another AI, one is a human. You've now read everything you need to tell them apart.