How to spot AI-written prose: the seven real tells

Most advice about spotting AI writing is folklore. People swear they can smell it — the em dashes, the word “delve,” the bullet points — but every one of those surface markers is trivially edited away, and plenty of humans use them naturally. Meanwhile the tells that actually survive editing are structural. They live in the shape of the paragraph, not the vocabulary, because the shape is what the machine is optimizing and the vocabulary is just what falls out of it.

What follows are the seven patterns we lean on most when writing the tells for the daily puzzle. None of them is proof on its own. Machine text is a probability cloud, not a fingerprint, and the honest way to use these tells is the way a doctor uses symptoms: one is a hint, three are a diagnosis.

1. Uniform polish

Human prose has a texture problem: some sentences are worked and some are just transportation. Writers polish the sentence they care about and let the others carry luggage. Machine prose polishes everything. Every sentence lands, every clause balances, every image is groomed — and the total effect, once you notice it, is of a lawn where every blade has been cut to the same height. When you read a paragraph and cannot find one boring sentence, be suspicious. Real writing wastes words the way real conversation wastes breath.

2. The parade of personification

One personification is a style. Four in a paragraph is a tic. Machine prose loves to make the world cooperative — heat has opinions, curtains decide, awnings drip in agreement — because each little animation scores well on its own. A human writer spends that move once and moves on, partly out of taste and partly because sustaining a whimsical universe is exhausting. The machine is never exhausted, and that is exactly what shows.

3. The gift-wrapped ending

Machine paragraphs are built backward from their best line. The closing sentence reframes everything — “It is the audience for it.” “It knew both songs.” “They were the only part of the country that never changed.” — and every sentence before it exists to set up the reveal. Humans bury their best line mid-paragraph all the time, because they wrote it when it occurred to them and didn't reorganize the furniture afterward. If the paragraph reads like a joke with a punchline, and the punchline is wisdom rather than humor, you are probably reading a machine.

4. Nothing is allowed to be boring

Read Mark Twain on a river town and you'll hit freight statistics. Read Thoreau on factory clothing and you'll wade through hedged qualifications — “as far as I have heard or observed” — before any aphorism arrives. Real writers include load-bearing dullness: dates, wages, measurements, the name of a street. Machine prose skips the boring parts because boring parts don't score. A passage where every detail is evocative and no detail is merely informative has the fingerprints of optimization all over it.

5. Symmetrical scaffolding

“Not the month, the one you carry.” “They lose the argument and keep the ceremony.” “We retired the horse and kept the horsepower.” The seesaw sentence — not X but Y, lose A keep B — is the single most reliable syntactic tell in contemporary machine prose. Humans use it too, but sparingly, because it draws attention to itself. When you see three seesaws in one passage, or a triad of examples so evenly weighted they feel like stock photography (clerks by candlelight, farmers at dawn, mothers with ledgers), the odds shift hard toward machine.

6. Emotion without embarrassment

Whitman addressed his own throat. Wharton let a grown man kiss a shoe. Dickinson rhymed “strategy” with “physiognomy” and dared you to object. Human writing at full power risks being ridiculous, and the risk is visible on the page. Machine writing is emotionally tasteful. It grieves gracefully, loves appropriately, and never once makes you cringe on its behalf. Perfect emotional manners are not a human trait; they are a training objective.

7. The missing wrong note

Woolf saw sheep grazing “on pointed wooden legs” and kept the perception error in. Twain coined “crackless” under pressure of a memory and never used it again. Dickinson wrote “say how many dew,” which isn't grammar at all. The deepest tell is the absence of these wrong notes — the moments where a human's perception outruns their language and something broken-but-true gets through. A machine can be prompted to fake brokenness, but faked brokenness is symmetrical too, if you look closely. Real damage has no style.

Using the tells honestly

Two warnings. First, these patterns describe default machine prose — text generated without much adversarial effort. A skilled operator prompting a model to write badly, or editing its output by hand, can defeat any single tell. Second, some humans write like machines: ad copy, corporate communications, and MFA-workshop fiction all optimize for the same judges the models were trained on, which is why they trip the same alarms. The game is probabilistic. Play it that way — and practice daily. Your ear improves faster than the models do, at least for now.