AIContent MarketingCommunicationsCopywriting

How to Spot AI-Written Content: 21 Tells Your Readers Are Already Noticing

By Laura Braden , VP of Communications

How to spot AI-written content — 21 tells your readers are already noticing

Originally written by Laura Braden.

Your readers may not be able to name what bothers them about AI-written content. They’ll just feel it. Something slightly off. A little too smooth. Too predictable. Too even. And then they’ll keep scrolling.

AI writing tools leave fingerprints. For anyone working in communications, public relations, or content marketing (fields where your judgment and voice are your professional currency), that’s actually useful information. The patterns are learnable, most are fixable, and recognizing them is now part of the job.

Here are 21 of the most recognizable AI writing tells, what they look like, and what to do instead. Not a case against using AI tools. It’s a case for using them with your eyes open.

Laura Braden is a senior communications strategist with more than 20 years of experience in nonprofit, civic, public health, and political communications. She serves as Chief Impact Officer at GoldenDoodle and leads PR and communications strategy at MKTNG.

The 21 tells

1. The AI vocabulary cluster

AI models dramatically overuse a recognizable set of formal, slightly literary words that almost never appear in natural professional writing. The current red-flag list: delve, tapestry, realm, beacon, nuanced, robust, multifaceted, intricate, cornerstone, harness, foster, illuminate, paramount, leverage (as a verb), and navigate (used metaphorically). Expect this list to shift as models evolve.

AI: “This initiative underscores the pivotal role of harnessing innovative strategies to navigate today’s multifaceted digital landscape.”

Human: “This project shows why we need new tactics for online marketing.”

The fix: Run a find-and-replace pass on these specific words before anything else. Each one has a simpler, more direct substitute one step behind it.

2. Significance inflation

AI inflates ordinary topics with grandiose language. A regional conference becomes “a groundbreaking testament to the transformative power of innovation.” A software update becomes “a paradigm shift in how organizations approach workflow optimization.” Intensifiers to watch for: pivotal, crucial, vital, revolutionary, game-changing, unprecedented, cutting-edge. When everything is extraordinary, nothing is.

The fix: Replace the inflation with the actual thing. What did the conference accomplish? What does the update actually change? Specific answers are more persuasive than inflated ones.

3. Em dash overload

By mid-2025, more than half of ChatGPT responses contained em dashes, up from fewer than 1 in 10 the year before. The Washington Post, NPR, and Rolling Stone all covered the phenomenon. Em dashes are legitimate punctuation. For the record, I genuinely mourn what AI has done to their reputation. But AI deploys them with mechanical regularity as a universal connector for emphasis, pivots, and parenthetical asides. Three or more in a single paragraph is a strong signal. Some editors have started calling it “the ChatGPT dash.”

The fix: Reread every em dash you didn’t consciously choose. Most can be replaced by a comma, a period, or a rewritten sentence that doesn’t need the connector at all.

4. The “It’s not X, it’s Y” construction

This is AI’s favorite rhetorical device. The negation-then-assertion pattern shows up constantly in LinkedIn posts, blog intros, and marketing copy: “We’re not just building a product, we’re creating an experience.” Two or more in a single piece are a clear red flag. AI often repeats the formula without registering it already used it. It’s a structure that sounds like insight without delivering any.

The fix: If you catch yourself writing “it’s not X, it’s Y,” ask what you’re actually trying to say. The real point is usually somewhere in the second half of the sentence. Start there.

5. The “In today’s world” opener

AI opens content with generic temporal framing that could apply to any topic in any decade. “In today’s rapidly evolving landscape” and its variants add zero information. Delete the first sentence of any piece that opens this way and read what came second. It will be stronger. It always is.

6. The rigid intro-body-conclusion structure

AI applies the same mechanical structure to every piece regardless of format: broad contextual opener, organized middle sections, tidy summary conclusion. Human writers vary structure based on purpose. Some pieces start mid-story. Some skip a formal conclusion entirely. When every piece from the same source follows the exact same arc, the machine origin shows.

The fix: Ask what structure this specific piece actually needs, not what structure you defaulted to.

7. Metronomic paragraph length

Researchers measure sentence variation as “burstiness,” and studies show AI text scores significantly lower than human writing on this metric. AI paragraphs tend toward uniform, predictable sentence length.

Human writing isn’t like that. Short punchy lines land a point. Longer ones unpack it. A one-sentence paragraph is fine. So is one that runs to eight sentences working through a single idea. The mix is the thing.

8. The rule-of-three default

AI consistently groups items in threes. Three examples. Three benefits. Three takeaways. Two feels incomplete to a language model; four feels excessive. Read back through a few AI-generated pieces and count the groupings. The pattern becomes obvious fast. The antidote is choosing how many examples a point actually needs, which is sometimes two and sometimes six.

9. Excessive headers and bold formatting

AI defaults to Title Case headings, dense subheaders, and a “Bold Term: explanation” inline format that rarely appears in organic professional writing. It treats every section as requiring a formal heading, even in short pieces. The result looks like a Wikipedia article when a memo would serve better.

The fix: Earn every header. If a section is two sentences long, it probably doesn’t need one.

10. Transition word overload

Researchers at Carnegie Mellon found that certain AI-favored transition words appear up to 150 times more often in AI outputs than in comparable human writing. “Furthermore” and “moreover” show the same pattern: technically correct, but not how anyone actually writes. When every paragraph opens with a formal connector, the accumulated effect is mechanical.

The fix: Read the piece aloud. If you wouldn’t say “furthermore” out loud, cut it.

11. “It’s important to note that…”

It sounds authoritative and adds nothing. Variants: “It’s worth noting,” “It should be noted that,” “It bears mentioning.” Delete every instance and read the sentence without it. It holds up every time. The phrase is pure throat-clearing.

12. Excessive hedging

AI systematically over-qualifies every claim. “Generally speaking, in many cases, content marketing can potentially be an effective strategy, though results may vary depending on a variety of contextual factors.” Research has confirmed that ChatGPT significantly increases hedges and reduces direct assertions compared to human writing. That is the exact opposite of persuasive professional writing.

The fix: Take a position. If you genuinely don’t know, say so directly: “I don’t know yet” is more honest and more readable than a paragraph of qualifications stacked on top of each other.

13. Missing contractions

AI defaults to “it is,” “you will,” and “do not” rather than “it’s,” “you’ll,” and “don’t,” even in contexts that call for warmth and a conversational register. Marketing copy, client emails, and social content written without contractions feel stiffly formal in ways that register even when readers can’t name them.

The fix: Do a contraction pass. If the piece calls for a warm, direct register, which most professional writing does, replace the formal constructions.

14. No real voice

AI text defaults to formal, emotionally detached prose. It can’t draw on real experiences, name specific colleagues, recall sensory detail, or express genuine surprise, frustration, or delight. It reads like a statistical average of a million voices: polished and weightless. Readers feel that absence even when they can’t articulate it.

This is the one no prompt fully fixes. It requires a real person with real opinions making deliberate choices on the page.

15. Locked grammatical person

AI holds a single grammatical person rigidly throughout a piece: entirely second-person “you,” or entirely third-person “they.” Human writers naturally shift between first, second, and third when telling stories, injecting personal reactions, or quoting sources. A 1,500-word article that never once steps outside second person carries a mechanical quality that experienced readers pick up on.

16. Hallucinated statistics

AI fabricates statistics, citations, case studies, and quotes and presents them with full confidence. One study of surveyed marketers found that nearly half reported completely false AI-generated information made it past review and went live publicly.

Editor note: Verify this stat before wider circulation. The original source in our draft was unattributed.

Every AI-generated data point needs independent verification before it reaches a client or a public audience. No exceptions.

The fix: Treat every AI-generated statistic as unverified until you’ve traced it to its actual source.

17. Vague attribution

AI cites authority without providing actual sources: “Research indicates,” “experts agree,” “a recent survey found.” This creates the appearance of evidence-based writing while providing nothing verifiable. Any claim that opens with one of those phrases and doesn’t end with an actual citation needs to be cut or sourced.

The fix: If you can’t find the real source, rewrite the claim as your own professional observation rather than borrowed authority.

18. Synonym cycling

AI avoids word repetition by cycling through synonyms unnecessarily. The same subject gets a new name every sentence: “The protagonist faces many challenges. The main character must overcome obstacles. The central figure eventually triumphs.” Pick one term and stay with it. Repeating a proper noun is not a writing error. Synonym cycling is.

19. Bullet point addiction

AI defaults immediately to numbered or bulleted lists, even for complex topics that deserve flowing prose. And it decorates those lists with emoji: checkmarks, brain icons, sparkles. Human writers use lists selectively, after prose has established context and earned the structure.

Asked “How can I be happier?”, a language model gives you five bullet points and a trophy emoji. A human writer gives you a paragraph that starts with an honest observation.

20. Defining terms the audience already knows

AI defines terms that professional readers don’t need defined. In a post written for marketing directors: “Search Engine Optimization, commonly known as SEO, is the process of improving your website’s visibility in search engine results pages.” Calibrate to who’s actually reading. Defining basic terms signals that the content wasn’t written with a specific audience in mind.

The fix: Ask “would a normal person in this role need this explained?” If the answer is no, cut the definition.

21. Each model has a distinct signature

ChatGPT leans heaviest on em dashes, bold formatting, bullet structures, and words like “certainly” and “utilize.” Claude writes more flowing prose with less formatting, favors measured hedging and introspective tone, and reaches for phrases like “it appears” and “it’s worth considering.” Gemini produces utilitarian, report-style output with more frequent italics and the construction “not only… but also.”

When multiple writers on a team use the same model without heavy editing, their output starts to converge into suspiciously uniform prose. That’s a team-level tell, and it’s increasingly recognizable to editors, journalists, and anyone who reads a lot of professional content.

The one thing no prompt can fix is a real voice. That still requires a real person. Your audience can tell the difference, even when they’re not trying to.

Frequently asked questions about AI writing tells

What are the most common signs that AI wrote a piece of content?

The most recognizable tells are overuse of a specific vocabulary cluster (delve, robust, nuanced, cornerstone, and related words), em dash overload, significance inflation of ordinary topics, rigid three-part structure, metronomic paragraph length, missing contractions, vague attribution without actual sources, and a formal detached tone with no genuine personal voice.

Can AI-written content be edited to sound human?

Most of the mechanical tells are fixable with a careful editing pass: replace the AI vocabulary cluster, vary paragraph and sentence length, remove vague attribution, add contractions, cut the “it’s important to note” filler. What’s harder to fix through editing alone is authentic voice and specific lived experience. That still has to come from a real person.

Do different AI models write differently?

Yes. ChatGPT uses em dashes heavily, leans on bold formatting and bullet structures, and defaults to words like “certainly” and “utilize.” Claude writes more flowing prose with less formatting and tends toward hedged, introspective language. Gemini produces more utilitarian, report-style output with phrases like “not only… but also.” All three share the core vocabulary cluster and structural tells listed above.

Why does AI writing feel off even when it’s grammatically correct?

Because it’s optimizing for a statistical average across millions of documents rather than for a specific reader in a specific context. It defaults to the most expected next word, which produces text that’s smooth but weightless. There’s no genuine experience behind it, no specific detail, no voice shaped by actual decisions and failures. Readers feel that absence even when they can’t name it.

Is it wrong to use AI writing tools?

No. The issue isn’t whether you use AI assistance; it’s whether the final work reflects genuine human judgment, your actual voice, and a real point of view. AI is a useful first-draft tool. It shouldn’t be the last hand on the work. The fix for off-brand AI drafts is upstream: give every tool the same record of your company. That’s what an AI knowledge base is for.

This post was originally authored by Laura Braden.

We use AI tools in our workflow where they add clarity and speed; this piece was reviewed by our team for accuracy and voice.

← All Posts