The Signs of AI Writing: 14 Tells I Look For in Every Draft
What Are the Word-Choice Signs of AI Writing?
Three of the fourteen signs live entirely in vocabulary, and I catch two of them from a single sentence before I’ve read the rest of the paragraph.
1. The Vocabulary Graveyard
A model reaches for the same handful of words at any given opportunity, because they’re statistically safe and they fit almost any sentence.
I’ve written the full graveyard list elsewhere, so here’s the short version: delve, leverage, tapestry, robust, seamless, elevate, unlock, navigate, landscape.
Flagged, from a coaching client’s about page: “We leverage a robust suite of tools to deliver a seamless client experience.”
What I sent back: “We use 11 tools to run your launch. You get one dashboard and one point of contact, not 11 logins.”
Same claim. One version has three graveyard words and nothing a reader can picture, and the other has a number.
2. The Cliché Starter Pack
Past the individual words, there’s a rotation of stock phrases a model reaches for no matter the topic. Think the fast-paced-world opener, the ever-changing-landscape line, and the embark-on-a-journey send-off.
Flagged, from a SaaS onboarding email: “As we embark on this journey together, we’re excited to help you navigate the ever-changing landscape of project management.”
That sentence could sit on any onboarding email for any product in any category, and nobody reading it would notice it was wrong for this one.
What I sent back: “You’ve got three tasks due this week. Here’s where to find them.”
3. Forced Sass Openers
This one’s sneakier, because it reads as personality instead of a tell.
A model throat-clears before a point with a short, punchy phrase that’s supposed to sound like a real person talking: “But real talk.” “The result?” “Hot take:” “And honestly?”
Flagged, from a founder’s LinkedIn draft: “Hot take: most agencies are lying to you about their results. The result? Clients who leave after three months.”
Two forced openers in two sentences, and neither one earned the punch it was reaching for.
What I sent back: “Most agencies pad their results, and the clients who catch on leave after three months.”
One sentence instead of two, and it says exactly the same thing without performing confidence it hasn’t demonstrated yet.
What Are the Sentence-Level Signs of AI Writing?
Five signs show up inside a single sentence once you know to look for the shape of it, not just the words.
4. The Em Dash Epidemic
A model uses the em dash as duct tape whenever two half-formed ideas don’t quite connect on their own.
Flagged, from a course-launch email: “This program delivers real transformation—and the results speak for themselves—students who’ve doubled their income.”
Three ideas welded together with dashes instead of picked apart into sentences that each carry their own weight.
What I sent back: “This isn’t a course you watch. It’s a course you do, and the students who finished it doubled their income within a year.”
I’ve written a full breakdown of how to make AI writing sound more human if the em dash is the only tell you’re hunting for, because it’s rarely the only one on the page.
5. Binary Contrast Addiction
The move here is to negate a plain claim, then assert a fancier-sounding version of the same idea.
Flagged, from a coaching client’s sales page: “Success isn’t about working harder, it’s about working smarter.”
That sentence has said nothing. It’s a shape borrowed from a motivational poster, not a claim about this specific coach’s method.
What I sent back: “You don’t need more hours. You need the two hours you already have pointed at the right client.”
6. Label-Colon Openers
A model announces the category of its point before making the point.
Flagged: “The truth: your pricing page is losing you money.”
What I sent back: “Your pricing page is losing you money.”
Nobody talks like the first version. A person doesn’t preview the type of sentence they’re about to say. They just say it.
7. The Metronome Sentence
This is the one that survives every synonym swap, because it’s not about the words at all.
AI defaults to medium-length sentences with almost no variation, usually somewhere around 12 to 18 words, one after another, at the same pace.
Flagged, from a nonprofit newsletter draft: “Our volunteers work hard every day. They serve meals to families in need. This impact ripples through the community. Every donation makes a real difference.”
Four sentences, four almost identical lengths, and not one of them sounds like a specific person wrote it.
What I sent back: “Rosa’s been on the serving line every Tuesday for six years. Last month she fed 340 families. That number only happens because someone kept showing up.”
Short, then medium, then a longer one that runs on a little, because that’s closer to how someone talks when they care about what they’re saying.
8. The Hedge Reflex
A model avoids committing to a claim by wrapping it in “might,” “could,” “perhaps,” “typically,” or “somewhat,” even when the claim is one it should just make.
Flagged, from a product update draft: “This update could potentially improve load times for some users, depending on their setup.”
What I sent back: “This update makes the dashboard load faster. Most people notice it inside the first click.”
The hedge isn’t caution. It’s the absence of a specific answer standing in for one.
What Are the Structural Signs of AI Writing?
Three signs stop showing up in one sentence and start showing up in how a paragraph, or a whole draft, is built.
9. The Perfect Triple
Three feels complete, and a model applies it everywhere whether three is the honest count or not.
Flagged, from a workshop landing page: “You’ll learn to plan, execute, and optimize your entire content strategy.”
What I sent back: “You’ll leave with a content calendar for the next 90 days. That’s it. That’s the workshop.”
Real thinking doesn’t produce a clean three every time. Sometimes there’s one thing that matters and four things that don’t.
10. Sentence Stacking
Every sentence in a flagged paragraph can be grammatically fine on its own, and the paragraph still reads like a printer instead of a person, because nothing connects one sentence to the next.
Flagged, from a B2B case study draft: “The client had a leaky funnel. Email open rates were declining. Landing pages weren’t converting. A new strategy was needed.”
Four facts dropped on a table with no mortar between them.
What I sent back: “The client’s funnel was leaking, and it showed up everywhere at once. Open rates were sliding, landing pages weren’t converting, so we needed a strategy that fixed both at the same time, not one after the other.”
I go deeper on this exact pattern, plus five others like it, in the sentence-rhythm breakdown, because sentence stacking alone accounts for more flagged drafts than any single word on the graveyard list.
11. The Suspiciously Tidy Draft
No typos. Every paragraph runs the same length. Every section lands on a neat, satisfying close, even the ones that shouldn’t have one yet.
Flagged, from a webinar recap: three paragraphs, each exactly four sentences, each closing on a version of “and that’s what made the difference.”
Real drafts are messier than that on the first pass. A person restructures a thought mid-paragraph, or leaves one section a sentence longer than the others because that section needed it.
What I sent back: I cut the third paragraph’s closing line entirely and let it end on the last fact instead of a bow.
Want this done for you? Book a free strategy call →
What Signs of AI Writing Only Show Up Across a Whole Page?
Three signs never show up in one sentence. You need the whole draft in front of you, sometimes the whole website, before the pattern lands.
12. The Empty Third Person
A flagged draft avoids “I” almost entirely, staying in second and third person even when a first-person moment would carry more weight.
Flagged, from a founder bio: “The founder built this company after seeing a gap in the market. Customers deserve better options.”
What I sent back: “I started this because I got quoted $40,000 for something I ended up building myself in a weekend. If I got burned that badly, I figured other people were too.”
The first version could describe any founder of any company. The second one could only be about the person who lived it.
13. The Voice That Resets Every Post
This one only shows up once you read three or four pieces from the same source back to back.
A model has no memory of the tone it used yesterday, so post one sounds breezy, post two sounds formal, and post three throws in jokes that never appeared before or after.
I caught this on a client’s blog where four posts, written across two months, each read like a different person had taken over. Once I pointed it out, they could see it too. They just hadn’t read all four in one sitting before.
14. The Confident Fake Fact
A model states a fabricated statistic, an invented citation, or a strangely precise number with the exact same confidence it uses for the true parts of the same paragraph.
Flagged, from a case study draft: “Studies show that 73% of B2B buyers abandon a purchase after one bad email.”
No study was cited, because no study said that. The number sounded precise enough that it almost slipped through.
I built a five-check pass specifically for this tell, because it’s the only one on this list that can get a client sued instead of just embarrassed.
I also tested 13 AI detectors against known samples, and most of them miss this exact failure mode, because a detector measures rhythm, not truth.
14 Signs of AI Writing at a Glance
| # | Sign | What it looks like | What I do instead |
|---|---|---|---|
| 1 | Vocabulary graveyard | Delve, leverage, tapestry, robust, seamless | Cut the word, name the real thing |
| 2 | Cliché starter pack | The fast-paced-world opener, the ever-changing-landscape line | Say the specific detail instead |
| 3 | Forced sass openers | ”Hot take:”, “The result?”, “And honestly?” | Say the point directly, no wind-up |
| 4 | Em dash epidemic | Two half-ideas welded together with a dash | Split into two full sentences |
| 5 | Binary contrast | Negates a plain claim, then reasserts a fancier version | State the real claim once |
| 6 | Label-colon openers | ”The truth:”, “Bottom line:“ | Delete the label, keep the insight |
| 7 | Metronome sentence | Every sentence 12-18 words, same pace | Short, then long, then short |
| 8 | Hedge reflex | ”Might,” “could,” “perhaps,” “typically” | Commit to the specific answer |
| 9 | Perfect triple | Everything grouped in exactly three | Vary the count on purpose |
| 10 | Sentence stacking | Facts listed with nothing connecting them | Add and, but, because, so |
| 11 | Suspiciously tidy draft | Zero typos, every section wraps in a bow | Leave one section rougher, on purpose |
| 12 | Empty third person | Avoids “I,” stays generic and safe | Ground the claim in one lived detail |
| 13 | Voice reset | Tone shifts post to post, no throughline | Read the last three pieces back to back first |
| 14 | Confident fake fact | A precise stat or citation nobody can source | Verify or delete before it ships |
How Do You Tell If Someone Is Writing With AI?
Read the paragraph out loud, and listen for whether the sentences connect to each other or just sit next to each other.
That single test catches more AI writing than any word list, because rhythm is harder to fake than vocabulary. A person can swap “leverage” for “use” in ten seconds. Fixing four flat, disconnected sentences takes actual rewriting.
If you want the fast version, run down the “what I do instead” column in the table above. Any sentence you can’t rewrite that way in under a minute probably wasn’t written by a person to begin with.
What Are Red Flags for AI Writing?
Four signs move faster than the rest, so I check these first on any draft that lands in my inbox.
-
Metronome sentence rhythm, because it shows up in the first paragraph whether the topic is marketing or medicine. 2. Vocabulary graveyard words clustered together, three or more in one paragraph.
-
Sentence stacking, where every fact sits next to the last one with no connector. 4. A suspiciously tidy draft, with zero typos and a bow tied on every section.
Any one of these alone could be a careful human writer having a stiff day. Two or three stacked in the same page is the signal.
How Do I Check If Writing Is AI?
Start with the read-aloud test, because it costs nothing and it’s more reliable than most of the paid alternatives.
If you want a second opinion, an AI detector can help, but go in knowing what they catch.
I ran 13 of the most popular detectors against three controlled writing samples on camera, and only one caught every sample correctly. Two of them flagged genuine human writing as AI-generated.
So treat a detector score as one data point. The read-aloud test plus the fourteen signs above will get you further than any single tool’s percentage.
What Words Do AI Use the Most?
Delve, leverage, tapestry, robust, seamless, elevate, unlock, navigate, and landscape sit at the top of the list, and I keep the full vocabulary graveyard updated with the newer entries as they show up.
One of these words in a long draft is invisible. Three or four in the same paragraph is a body count.
What These Signs Cost You If You Miss Them
A reader scanning your page isn’t grading your grammar. They’re deciding, inside the first few seconds, whether a real person who knows the business wrote it.
I watched this happen on a sales call a coaching client almost lost. The prospect told her, before she even pitched, that her website “read like ChatGPT wrote it,” and he’d nearly closed the tab before she got him on the phone.
Nothing on that page was factually wrong. It just carried enough of these fourteen signs stacked together that a stranger caught it in under ten seconds, without knowing a single term on this list.
That’s the stakes. Forget the detector score and the word count. It comes down to whether the next person reading your page believes a specific human being wrote it for them.
I run this same fourteen-point check, plus a broader 25-point editing pass, against every draft that comes through my inbox before it ships to a client.
If you’d rather run it yourself first, the free copy cleaner checks a draft against most of these fourteen signs in one pass, the same check I used to do by hand.
FAQ
How do you tell if someone is writing with AI?
Read a paragraph out loud and listen for flat, same-length sentences with nothing connecting them. That single check catches more AI writing than any word list, because rhythm is harder to fake than vocabulary.
What are red flags for AI writing?
The fastest four to spot are a metronome-flat sentence rhythm, disconnected sentences with no logic between them, a suspiciously tidy draft with zero typos, and vocabulary like delve, leverage, or seamless clustered in one paragraph.
How do I check if writing is AI?
Start by reading it out loud, then run it through a detector if you want a second opinion, knowing detectors miss more than people expect. I tested 13 of them against known samples and only one caught every sample correctly.
What words do AI use the most?
Delve, leverage, tapestry, robust, seamless, elevate, unlock, navigate, and landscape top the list. One of these in a long draft is invisible. Three or four stacked in one paragraph is the tell.
What are some common AI writing tropes?
Beyond vocabulary, the tropes I flag most are the forced rule of three, the it's-not-X-it's-Y contrast, the label before the insight, and a closing paragraph that wraps everything in a neat bow no real draft earns on the first try.
Is there a name for the AI writing style?
Not one official name, but editors and detection researchers describe it the same way: high on polish, low on burstiness, meaning the sentence lengths barely vary and nothing sounds like a specific person said it.
Can these signs show up in a human's writing too?
Yes, especially the hedge reflex and the forced triple, since both were rhetorical habits long before language models existed. The tell isn't one pattern appearing once. It's four or five of them stacking in the same page.
Want this built for your business? Book a free strategy call →