How to Tell If Text Is AI-Written: 12 Signs Editors Look For

No single word proves a text is AI-written: not "delve," not "tapestry," not the em dash that half of LinkedIn now treats as a confession. Editors who catch machine prose reliably aren't hunting a smoking gun; they're counting small uniformities that pile up. One tell means nothing, and a dozen sustained across three pages is a different story.

How to tell if text is AI written, by hand: look for uniform sentence rhythm, same-shaped paragraphs, hedged both-sides framing, generic examples, missing lived detail, and confident statements that turn out to be wrong. One sign means nothing. Four or five together, sustained across a whole piece, is a strong signal worth confirming with a detector run.

Why editors still catch what software misses

In that 2025 ACL study, expert annotators who frequently use ChatGPT identified AI-generated text with 99.3% accuracy and stayed accurate against paraphrasing tricks that defeat automated tools. Wikipedia has institutionalized the same skill: volunteers in its WikiProject AI Cleanup maintain a working catalog of machine-writing patterns they use to find generated text in articles. None of these people run statistics in their heads. They pattern-match on the signs below.

There's a supply-side explanation too. Editors, teachers and moderators now read machine prose all day, involuntarily, and repetition trains the ear the way any dialect does. The 99.3% figure came from exactly that population: people marinated in the style.

Machines read entirely different evidence, measuring how statistically predictable each word is given the ones before it; the mechanics live in our guides to how AI detectors work and perplexity and burstiness. The practical upshot: your eye and a scanner fail on different texts, which is exactly why you want both.

How to tell if text is AI written: signs 1–6

The first six signs are structural. Structure is where models leak, because it's the habit they can't stop performing.

A note on method before the list: read the whole piece once for meaning, then go sign-hunting on the second pass, and tally what you find. One or two hits describe most careful writing on earth. Four or five sustained hits, spread across different categories, are when suspicion earns its keep.

1. Metronome sentences. Every sentence lands between 15 and 25 words, forever. Human writing sputters and sprawls. If reading it aloud feels like a rowing machine, mark it.

2. Paragraphs stamped from one mold. Topic sentence, two supports, tidy mini-conclusion, repeat. Five paragraphs of identical build is a stronger tell than any word choice.

3. Both-sides hedging. "While X offers significant benefits, it is essential to consider the potential drawbacks." Nothing is claimed, nothing is risked, and the prose refuses to have a stake in its own argument.

4. The rule-of-three reflex. Everything arrives in triplets: "fast, scalable, and secure," "students, educators, and institutions." Humans use triads for emphasis; models use them as wallpaper.

5. Inflated vocabulary. "Delve," "underscore," "pivotal," "multifaceted," "rich tapestry." No word convicts on its own, and these lists rot as models update, but a pileup of conference-keynote diction in a casual context earns a margin note.

6. Punctuation that performs. Em dashes and colons carrying the rhythm of every single paragraph. Yes, human writers use them too, including us; the tell is density and sameness, not existence.

Signs 7 through 12: content and voice

7. Examples from nowhere. "Consider Sarah, a small business owner looking to grow her online presence." No town, no product, no price, no year. Real examples have grit and consequences.

8. No lived detail. Nothing that could date the piece, embarrass the author, or be checked against reality. Human writing leaks biography constantly; machine writing is biographically sterile.

9. Confident wrongness. A fabricated statistic delivered smoothly, a citation to a paper that doesn't exist, a quote nobody said. It usually looks like "a 2022 Harvard study found that 73% of consumers..." where no such study exists. This is the most damaging tell on the list, so spot-check one verifiable claim before you trust the other ten.

10. The summary sandwich. An opening that promises to explore the topic, a closing that begins "In conclusion" and restates every point in order. Efficient, symmetrical, and weirdly proud of itself.

11. Grammar too clean, zero risk. No fragments. No slang, no parenthetical aside that made you smile, no sentence breaking a rule on purpose. Perfect compliance reads as machine because humans buy style with small errors.

12. It answers a slightly different question. Ask why your Q3 churn spiked and receive "factors that can influence customer churn." Drift toward the generic version of the prompt is pure model behavior, and it survives even careful editing.

Two habits missed the numbered list but deserve honorable mention: an allergy to the first person singular, and transitions doing ceremonial work ("Having examined X, we now turn to Y") in a piece with no we and no turning. Neither proves anything alone. By now you know the refrain.

What these signs can't do

Every sign above also occurs in honest human writing. Formal genres reward hedging. Second-language writers are often taught the five-paragraph mold explicitly. Corporate style guides mandate triads and forbid fragments. Automated detectors share the same blind spot, and it has measurable victims: a Stanford study found detectors flagged an average of 61.3% of TOEFL essays by non-native English speakers. Human sign-reading inherits that bias unless you consciously correct for genre and background, a problem we unpack in why human writing gets flagged.

The list also ages. Signs 5 and 6 are already weaker than they were in 2023, because model vendors read the same complaints everyone else does and train the tics away. The structural signs decay slower: rhythm, paragraph architecture and prompt drift survive model generations because they come from how these systems compose, not from any word list. Weight your tally accordingly.

Confirming your read with a scan

Your eye and a detector disagree productively, because they're reading different layers of the same text:

The machine side has come a long way since 2019, when the GLTR project first colored words by how predictable they were. AI Detector 360 now runs text through multiple engines, paints a sentence-level heatmap of what looks generated, and labels the confidence of the whole read; the free AI detector does this on up to 5,000 characters without a sign-up, and how we calibrate those scores is documented on our methodology page.

Disagreement between the two is informative in both directions. A text your eye clears but the scanner flags is often formal human writing, the classic false-positive pattern. A text your eye flags but the scanner clears is often edited AI, where paraphrasing scrubbed the statistics and left the sterile content behind. Each combination tells you which evidence to gather next.

The routine that works: read first and mark the signs, then scan. If both point the same direction, you have grounds to go looking at drafts and history. If they disagree, believe neither yet. That's not indecision; that's the method.