Let’s skip the theory for a moment and do something more honest. Below are three pairs of paragraphs, each pair on the same topic. One of each pair is written the way an AI assistant tends to write when left to its own patterns. The other is written the way a person usually writes. Read each pair, pick your answers, and then we’ll go through what the differences actually are.
Pair 1 — Mornings:
A. Mornings are a universal battleground. The alarm rings, the day begins, and millions struggle to rise. It’s not about laziness. It’s about the collision between a body that wants rest and a schedule that demands motion. To master the morning is to master the day. Small habits matter: water, sunlight, a quiet moment before the noise begins. Consistency, not intensity, is the real key. Those who embrace the routine find that mornings stop being an obstacle and start becoming an opportunity.
B. I’ve hated mornings my whole life, which feels embarrassing to admit as an adult. I’ve read every productivity blog promising the 5 a.m. miracle, and they all skip the part where you’re awake at 2 a.m. wondering why you’re doing this. The one thing that actually helped was ridiculous: I stopped setting my alarm to a number ending in zero. Some days I’m still late. The 5 a.m. people would hate me, and honestly, that’s fine.
Pair 2 — Why people quit:
A. In today’s dynamic workplace, employees increasingly choose to leave organizations that fail to prioritize their well-being. Companies must recognize that engagement is not merely a buzzword but a fundamental driver of retention. By fostering a culture of transparency, offering meaningful growth opportunities, and demonstrating genuine appreciation, leaders can navigate the complex landscape of modern employment and unlock the full potential of their teams.
B. Jake put in his notice on a Tuesday, three weeks after his manager took credit for his launch work in a meeting Jake wasn’t invited to. It wasn’t the salary. It wasn’t the hours — those were mostly fine. It was that he’d stopped being able to describe what he did all day without feeling embarrassed. Nobody in HR ever found out. The exit interview got a polite answer. The real reason went home with him.
Pair 3 — Phones in the morning:
A. Recent studies show that a staggering 87.4% of young professionals check their smartphones within the first five minutes of waking, a 12.3% increase from the previous year. Experts believe this behavior significantly impacts productivity, with research suggesting morning screen time can reduce focus by up to 31.2%. As technology evolves, these numbers are only expected to grow, making it imperative for individuals to reassess their morning routines.
B. Somewhere I read that most people look at their phones within minutes of waking up. I don’t remember the exact number, and I suspect whoever published it pulled it from a survey of undergrads in one country. What I can tell you from watching my own household: it happens fast, it’s rarely about anything important, and the first notification usually isn’t worth the ten minutes it costs. I’d like to say I’ve fixed this. I have not.
The reveal
All three A samples are written in the patterns typical of AI output. All three B samples are written with the texture of a person. If you guessed all six correctly, you’re reading well — but the point of this exercise isn’t to test you. It’s to show you which specific differences you were picking up on, because they’re almost never the ones people assume.
Signal one: the rhythm never breaks
Read sample A1 again and listen to its pace. Every sentence is roughly the same length. Each one lands, then the next one lands, like a metronome. There are no interruptions, no digressions, no sentences that start and then turn. The paragraph builds to a neat, summarizing close — “mornings stop being an obstacle and start becoming an opportunity” — because the whole structure was designed to arrive there.
Sample B1 is lumpy. It has a short sentence, then a long one, then a confession (“which feels embarrassing to admit as an adult”), then a joke that isn’t trying to be profound (“Some days I’m still late”). Human writing has this bumpy, self-interrupting quality. AI output, unless someone forces it otherwise, settles into a smooth, even cadence — and smoothness across a whole paragraph is a real tell. Not because smooth writing is impossible for humans, but because the uninterrupted smoothness is rare.
Signal two: every sentence wants to be quoted
Here’s a test for any paragraph: count the sentences that could be pulled out and shared as an inspirational quote. A1 has three or four of them. “Consistency, not intensity, is the real key.” “To master the morning is to master the day.” That’s aphorism density — the habit of making every line land like a fortune cookie. Real writers produce these occasionally, as a treat. AI produces them as a default, because they sound like strong writing and the model has learned that strong writing is what people want.
There’s a related construction that shows up constantly and is worth learning to spot: the “It’s not X. It’s Y.” inversion. “It’s not about laziness. It’s about the collision between a body that wants rest and a schedule that demands motion.” The pattern feels insightful — it sets up a wrong assumption, then corrects it — which is exactly why AI loves it. It’s a great device. It’s also one of the most consistent fingerprints of unedited AI prose. Once you see it, you’ll see it everywhere, including in the writing of people who are just heavy users of AI drafting.
Signal three: it refuses to commit to a detail
Sample A2 is the giveaway of the whole set. Read what it actually says. “Employees.” “Organizations.” “Leaders.” “Culture.” “Transparency.” Every noun is a category. Nothing specific happens in that paragraph — no person, no date, no cost, no conversation, no consequence. It could describe any workplace, which is its point: it’s engineered to be safe and generic.
B2 is the opposite. It commits immediately — “Jake,” “a Tuesday,” “three weeks after his manager took credit.” The commitment is what makes it feel human, because a person writing about a real event has no choice but to include details. They’re describing something that happened. An AI describing a category of things has no details to give, so it stays abstract. This is the strongest signal in the whole guide: specificity is expensive to fake. The abstract, category-level noun — the “stakeholders,” the “journey,” the “ecosystem” — doing all the work is the surest sign that no human experience is behind the words.
Signal four: precision that’s too round to be real
Sample A3 shows the second-strongest signal. “87.4%.” “12.3% increase.” “Up to 31.2%.” The numbers aren’t just present — they’re absurdly precise, which is the tell. Real statistics in articles get rounded (“nearly nine in ten,” “roughly a third”), because real survey data doesn’t land on 87.4 percent unless it was sampled carefully. And the sources are ghosts: “recent studies,” “experts believe,” “research suggesting” — never named, never dated, never linked. B3 does the human thing: it admits it doesn’t remember the number, questions where it came from, and shares a specific observation instead.
This signal is also the easiest one to confirm, which matters, because style is only ever an indicator. Round-sounding precise numbers can be checked in thirty seconds. When an article says “up to 31.2%,” search for the study. It won’t exist, and you’ll have your answer.
What looks like a signal but isn’t
Now the part everyone gets wrong. The signs above are probabilistic — they raise the likelihood, they don’t prove anything. And some famous “signals” aren’t signals at all.
The em dash. This is the one everyone’s heard, and the title of this guide is a joke at its expense. Yes, AI writing loves em dashes. But so do a lot of excellent human writers — this sentence has one. And more importantly, em dashes are trivially easy to remove, so their absence proves nothing and their presence proves nothing. Treat them as noise.
The vocabulary itself. “Delve,” “landscape,” “robust,” “unlock,” “furthermore,” “in today’s fast-paced world.” These words do correlate with AI output. But they also correlate with mediocre corporate writing from 2009, and any AI output has likely been edited by someone who knows the clichés. A single flagged word is worthless. What matters is the density — if you hit three of these in a paragraph, start reading more carefully.
AI detector software. This deserves its own warning. Commercial AI detectors are demonstrably unreliable: they flag original human writing as AI, miss actual AI writing that’s been lightly edited, and their error rates shift as models change. If you treat a detector’s score as proof, you will eventually accuse a real person of something they didn’t do — a documented failure mode with real human costs. Use them as a nudge toward a closer read, never as a verdict.
And the big one: any of these style signals can be removed. A person who writes with AI and edits carefully, or who prompts the model to avoid its own patterns, produces text that passes every stylistic test. The signals in this guide are a spotlight, not a wall. AI content is getting harder to distinguish from the real thing by design, and style is the layer being erased first.
The signals that actually carry weight
For quick reference, here’s what to weigh when you read something suspicious:
| Signal | What it looks like | Reliability |
|---|---|---|
| Even, unbroken rhythm | Metronome cadence, no digressions | Moderate — easy to notice, easy to miss |
| Aphorism density | Too many quotable lines, fortune-cookie prose | Moderate |
| The “It’s not X. It’s Y.” move | Set up a wrong idea, correct it, repeat | Low-moderate — great device, overused |
| Category-level nouns | Stakeholders, culture, leaders, no names or dates | High — hardest to fake |
| Ghost precision | “87.4%,” “experts believe,” no source | High — fastest to verify |
| No self-correction | No “wait, actually,” no reconsidering | Moderate |
| No edge cases | Everything is universally true | Moderate-high |
How to confirm instead of guess
Style reading gets you to “this is worth checking.” Getting to an answer requires the boring step: verification.
- Check one fact. Find the most specific claim and chase it for thirty seconds. Fake precision collapses instantly.
- Look for the committed detail. Does the text name a person, a place, a date, a price? If a piece about work never mentions a single actual work event, treat it with suspicion — it’s describing a category, not an experience.
- Ask the source questions. If it’s signed, follow up. “Where did the 31.2% come from?” A person can answer. An author who’s never going to respond to you — or doesn’t exist — can’t.
- Hold both possibilities. A text can be human and bad. A text can be AI and accurate. The goal isn’t to catch machines; it’s to decide how much to believe. Likelihood is the honest currency, not certainty.
The honest verdict
Here’s the uncomfortable truth at the bottom of all this: you cannot reliably detect AI-written text from style alone, because the style can be cleaned, and the tools keep getting better at sounding like people — and people are getting better at noticing, which is exactly why the clean-up matters. The signals in this guide raise and lower your confidence. They do not give you a verdict.
What that means practically is that the real defense against machine prose isn’t reading style closer — it’s reading content harder. Verify the facts. Look for the committed details. Ask where things came from. The signals we walked through are a good first filter, and spotting them genuinely helps. But the moment the filters stop working — and for well-edited text, that moment has already arrived — the only thing left that protects you is the habit of checking what you’re being told, regardless of who or what typed it.
That habit is the one signal that never fakes.
Independent technology writer focused on artificial intelligence, emerging technologies, and digital innovation. Covers AI applications in sports, productivity, and online business.
