Signs of AI writing: the complete list of tells

These survive a full vocabulary rewrite, which is why they are the real list. A model instructed to avoid "delve" will comply and still write prose with every one of these.

• Uniform paragraph lengthEvery paragraph lands at three to five sentences. Humans write one-sentence paragraphs when they are making a point.
• Uniform sentence lengthLow variance in sentence length is the strongest statistical signature of generated prose. Mix a four-word sentence in.
• The rule of threeEverything arrives in threes: three adjectives, three bullets, three reasons. Once is rhetoric, five times is a template.
• Bold lead-ins in every bulletA bold phrase followed by an explanation, repeated down a list, is model formatting.
• Header every two paragraphsReal articles let an argument run. Model output chops everything into scannable chunks.
• Question headingsH2s phrased as questions the body then answers in order — useful in SEO copy, robotic when every heading does it.
• The summary paragraphEnding with a paragraph that restates the piece is a model habit. End on the last real point.
• Both-sides hedging"While X has benefits, Y also has drawbacks." Real writers take a position.
• Zero specific detailNo dates, no names, no numbers, no place. Generation is the art of staying unfalsifiable.
• The flattering openerA first line that praises the question before answering it.
• Restating the promptThe first sentence repeats what was asked instead of answering it.
• Symmetrical listsEvery item in a list is the same grammatical shape and almost the same length.
• No digressionsHuman writing contains something irrelevant. Model output never wanders.
• Consistent registerNo shift between formal and casual, no aside, no joke that might not land.
• Emoji section markersEmoji used as headings rather than in place of a reaction.
• Tables where prose belongsA three-column table for something that is really a sentence.
• Titles in Title CaseMixed conventions, or Title Case applied to a sentence-style heading set.
• Zero contractionsIn casual contexts, writing "it is" everywhere instead of "it's" reads generated.
• The disclaimerA closing reminder to consult a professional, whether or not one is relevant.
• Over-signposting"First, let's look at… Next, we'll cover…" Narrating structure instead of using it.

Word-level signals are the ones people talk about most and should trust least. Individual words drift between model versions, and a subject-matter expert legitimately writes about realms, landscapes and intricate systems. Density is what carries information, and the full word list is maintained separately so it can be kept current.

Punctuation is the most visible layer and the weakest evidence, so this page is the one place in the cluster that does not print the full character table. The short version: the spaced em dash is the only punctuation tell with real signal, and only at several per paragraph; typed --, arrows, section signs and check marks are weak; curly quotes and ellipses are correct typography and not tells at all. The full table with per-character notes is maintained in exactly two places: the punctuation reference and the word and punctuation list, both driven from the same catalogue — printing it a third time here would just give you three copies to go stale.

• A single em dash. Standard punctuation since the eighteenth century. See do em dashes mean AI for the fuller argument.
• Curly quotes and curly apostrophes. Word and Google Docs convert these automatically. If it passed through a word processor, it has them.
• Perfect grammar. Copy editors exist. So do non-native speakers who write more carefully than most native ones.
• Formal or corporate register. Lawyers, consultants and academics wrote like that before models did.
• Short sentences and simple vocabulary. This is the describe-what-you-see school of writing, and it is the opposite of a tell.
• A detector percentage. Detector scores measure token predictability, and they disagree with each other constantly. See how accurate AI detectors are.

Most tell lists are folklore that was never measured, so here is a measurement you can check. We ran the site’s own checker over a paragraph written in full AI-house style — catalogue words, even sentences, a dash — and then over the 13 edited posts on this blog, using the counting code that ships in the checker. The point is not the absolute numbers; it is what moves and what does not.

Two findings worth internalising. First, sentence spread separates the columns more cleanly than any vocabulary count: the generated paragraph never leaves an 11-to-18-word band, while edited prose swings from 4 words to 74. Second, dash density overlaps completely — the edited post about punctuation is denser than the machine-style paragraph, because it quotes dashes. A number that a careful human writer trips over is a warning, not a verdict.

1. Count the concrete details: dates, figures, names, places, quotes. Below one per 150 words is a warning sign.
2. Measure the sentence lengths. If the shortest sentence is within a few words of the longest across twenty sentences, the rhythm is machine-set.
3. Check the ending. A final paragraph that restates the piece is a model habit rather than a human one.
4. Run the counts. The local checker reports dash density, catalogue hits, sentence-length spread and structural flags from a single paste, and nothing leaves the browser.

Then stop. The honest conclusion from a list of tells is "this has the habits", never "this person used AI". Those are different claims, and only one of them is supportable from text alone.