Make AI sound human.

AI is a decent writer with one flaw. It sounds like everyone at once. The same phrases, the same sentence shapes, the same suspiciously tidy rhythm, across every tool and every topic. This page collects the tells that give machine writing away and ends with a prompt you can paste into ChatGPT, Claude, or Gemini to strip them out. The models change every few months, so we keep adding to it.

Why AI sounds like that

A language model writes by predicting the most likely next word. Most likely means most common, and most common means average. Trained on a huge slice of the internet, its default voice is the internet's average voice: a LinkedIn post, a press release, and a marketing blog blended smooth.

The second layer comes from how models get polished after training. They're tuned on human ratings, and raters reward answers that feel complete, balanced, and confident. That tuning is why AI hedges ("it's worth considering"), why it formats everything into tidy bullets of three, and why it reaches for the same dramatic pivots over and over. The output is grammatically perfect and completely without fingerprints. Grammarly's own analysis says the same thing. Models gravitate to whatever appears most often in their training data, so the most statistically safe phrasing wins every time.

AI remains a strong first-draft engine through all of this. It just writes with a generic accent, and the accent is removable. You remove it by telling the model, specifically, what to avoid.

The sentence patterns

Words get most of the attention, but sentence shapes are the bigger tell. The most famous one is the contrast pivot, and once you see it you can't stop seeing it:

• "It's not just X, it's Y."And the whole family: "It's not about X, it's about Y," "You don't need X, you need Y," "Not because X, but because Y," "X is no longer Y, it's Z."
• "No X. No Y. Just Z." The staccato triple. Punchy the first time, a tell by the hundredth.
• "Here's the thing."With its cousins: "Here's what's happening," "Here's the reality," "Here's what I'm seeing."
• "The problem? The result? The catch?" One-word cliffhanger questions, colon openers like "The takeaway:", and the dramatic mid-sentence setup ("It comes down to one thing: trust."). All of them stand in for writing the actual sentence.
• "One thing kept repeating…"Fake field observations: "What stood out?" "I've seen this cycle repeat," "Ever notice how…" The model has observed nothing. It's borrowing the voice of someone who has.
• "Because honestly? Look, I get it." Manufactured casualness, usually right before a sales pitch.
• "The quiet shift." Everything is quiet or happening quietly: quiet growth, quiet revolutions, quiet confidence, quietly taking sales. An odd quirk, and one of the easiest to catch.
• Two-word punch phrases, repeated."Short sentences. Big claims." Used two or three times per post for rhythm, and the rhythm is always the same.

The rule of three deserves its own paragraph. AI loves lists of exactly three, sentences with three parallel clauses, and paragraphs of three sentences. People vary. Machines don't.

The words

No single word proves anything, and quoting a source is always fine. The tell is frequency. When three of these show up in one paragraph, readers start hearing the machine.

The 27 rules

This is the working list we apply to our own writing and hand to clients. It overlaps with the patterns above on purpose. The patterns explain what to watch for; the rules are phrased so an AI can follow them.

1. Skip the "not X, but Y" pivot. If a sentence starts by saying what something is not, rewrite it to say what it is. AI leans on the negation setup constantly.
2. Watch the em dashes. One per piece at most. Periods, commas, and parentheses cover the rest.
3. No "but here's the thing." Go straight to the point it was teasing.
4. Stop attaching "quiet" to everything. No quiet shifts, quiet revolutions, or quiet confidence, and nothing happens "quietly." Once you notice it, it's everywhere.
5. Drop "That said." And "Having said that." Start the next sentence on its own content instead of a throat-clearing transition.
6. No cliffhanger questions. "The catch?" "The kicker?" "The problem?" A one-word teaser before a reveal. Put the reveal in a full sentence.
7. Don't lean on "represents" or "reflects." Or "embodies" or "stands for." Symbolic, vague language. Say concretely what the thing does.
8. Watch the colons. No label openers ("The takeaway:") and no dramatic setups ("Every job came down to one thing: access."). Write the sentence. A colon before a genuine list is fine.
9. Avoid the "noise" tic. One of AI's most overused words, usually inside "cut through the noise." Skip the metaphor.
10. Don't say "load-bearing." The word is now a tell on its own.
11. No vague "the ones who." Name the group: line cooks, solo bookkeepers, HVAC techs.
12. Mind expletive openers. "It is important to note that…" and "There are several reasons…" bury the subject. Put the real subject first.
13. No fake-universal openers. "In today's fast-paced world" opens nothing. Start with a specific fact, example, or claim.
14. Mind "Furthermore." Also "Moreover" and "Additionally." Connect ideas through their content, or just start the next sentence.
15. Vary how each sentence begins. No three in a row starting with the same word or the same structure.
16. Don't stack "whether" or "from X to Y." Once each per piece, not three times a paragraph.
17. Use contractions. Don't, it's, you'll. Spelled-out forms read stiff; save them for emphasis.
18. Mix sentence lengths. Short sentences next to long ones read human. Three medium sentences in a row read machine.
19. Vary sentence shapes. Not every sentence should balance around a comma ("X happened, and Y followed"). Simple sentences and the occasional fragment help.
20. Cap the "-ing" verbs. Empowering, enabling, showcasing, highlighting. At most one per paragraph.
21. Don't bullet what isn't a list. Prose by default. Bullets are for genuinely enumerable things, not explanations.
22. No invented named examples. No "Sarah Chen" or "Acme Corp." Use anonymized real references: "a roofer we worked with in 2025."
23. Use specific examples. A place, a year, a role. "A Durango boutique in 2026," not "a business owner."
24. Skip fake relatability. No "soft morning light," "just me and my coffee," or "we've all been there."
25. No soft hedging. Cut "it's worth considering" and "some might argue." Take a position and back it.
26. Cite and date numbers. Every statistic needs a source and a year, or it gets cut.
27. Vary list lengths. Use 2, 4, or 5 items. AI defaults to exactly three, every time.

The paste-in prompt

Everything above, written out as instructions an AI can follow. Paste it before your request, or drop it into a Project or custom instructions so it rides along with every chat. It works in ChatGPT, Claude, and Gemini.

Write in plain language, in my voice. The rules below target patterns that are common in AI writing and make text read as machine-made. Follow all of them.

Sentence patterns to avoid:

- Contrast pivots. Don't build sentences that negate one thing to assert another: "It's not just X, it's Y." "It's not about X, it's about Y." "You don't need X, you need Y." "Not because X, but because Y." "X is no longer Y, it's Z." "No X. No Y. Just Z." Say what something is, directly, without the negation setup.
- Setup fragments. Don't open with "Here's the thing," "Here's the reality:", "Here's what I'm seeing:", or "Hot take:". Start with the actual point.
- One-word cliffhangers. Don't write "The catch?", "The kicker?", "The problem?", or "The result?" as standalone teasers before a reveal. Put the information in a full sentence.
- Colon setups. Don't start sentences with a label and a colon ("The takeaway:", "The lesson:"), and don't build drama with a mid-sentence colon ("It comes down to one thing: trust."). Write it as a normal sentence. A colon that introduces a genuine list is fine.
- Fake observations. Don't claim to have noticed things: "One thing kept repeating," "What stood out?", "Ever notice how...", "I've seen this cycle repeat." Only report observations I actually gave you.
- Faux-casual asides. Don't write "Because honestly?" or "Look, I get it."
- Repeated punch fragments. Don't use two-word sentence fragments for rhythm more than once per piece ("Short sentences. Big claims.").

Style rules:

1. Make points directly. If a sentence starts by saying what something is not, rewrite it to say what it is.
2. Use at most one em dash in the whole piece. Periods, commas, and parentheses cover the rest.
3. Never write "but here's the thing." Go straight to the point it was teasing.
4. Don't use "quiet" or "quietly" as modifiers. No quiet shifts, quiet revolutions, quiet confidence, or "quietly taking sales."
5. Don't use "That said" or "Having said that" as transitions. Start the next sentence on its own content.
6. Don't use "represents," "reflects," "embodies," or "stands for" to describe significance. Say concretely what the thing does.
7. Don't use "noise" or "signal" metaphorically, as in "cut through the noise."
8. Don't use the word "load-bearing."
9. Name groups specifically. Instead of "the ones who show up every day," name them: line cooks, solo bookkeepers, HVAC techs.
10. Avoid empty-subject openers like "It is important to note that..." or "There are several reasons...". Put the real subject first.
11. No sweeping openers like "In today's fast-paced world" or "In an era of...". Open with a specific fact, example, or claim.
12. Don't use "Furthermore," "Moreover," or "Additionally." Connect ideas through their content, or just start the next sentence.
13. Vary sentence openings. Never start three sentences in a row with the same word or the same structure.
14. Use "whether" at most once per piece, and "from X to Y" at most once per piece.
15. Write with contractions (don't, it's, you'll). Spell words out only for deliberate emphasis.
16. Mix sentence lengths. Put short sentences (under 8 words) next to long ones (20 or more). Don't write three medium-length sentences in a row.
17. Vary sentence shapes. Don't let every sentence balance around a comma ("X happened, and Y followed"). Use simple sentences and the occasional fragment.
18. Limit "-ing" verb constructions like "empowering," "enabling," "showcasing," "highlighting." At most one per paragraph.
19. Write prose by default. Use bullet points only when the content is genuinely a list, never to format an explanation.
20. Don't invent named people or companies as examples (no "Sarah Chen" or "Acme Corp"). Use anonymized real-sounding references: "a two-location dental practice," "a roofer we worked with in 2025."
21. Make examples specific, with a place, a year, or a role. Write "a Durango boutique in 2026," not "a business owner."
22. No fake sensory relatability. No "soft morning light," "just me and my coffee," or "we've all been there."
23. No hedging. Cut "it's worth considering," "some might argue," and "perhaps it's the case." Take a clear position and back it.
24. Only include numbers you can attribute. Every statistic needs a source and a year, or it gets cut.
25. Vary list lengths. If the piece has more than one list, don't make them all three items long.

Go easy on these words (once is fine, a pattern is a tell): gentle, quiet, quietly, noise, precise, furthermore, load-bearing, that said, honestly, testament, underscore, unwavering, foster, empower, amplify, leverage, harness, pivotal, profound, seamless, journey, delve, robust, cutting-edge, transformative, elevate, unlock, unleash, landscape, tapestry, realm, beacon, roadmap, game-changing.

If following a rule makes a sentence stiff or unnatural, sounding natural wins.

Two things the list can't do

A banned-word list improves a draft; it doesn't supply a voice. The stronger move is pairing the don'ts with two or three short samples of writing you actually like, yours or anyone's, and telling the AI to match them. Positive examples beat prohibitions.

And the list ages. Every model release retires some tells and mints new ones. "Delve" was the giveaway of 2024; "quiet" and "noise" took over later. When you spot a phrase we're missing, email it to us and we'll add it.

Sources and further reading

Pieces we checked this list against, worth a read if you want to go deeper: Grammarly on common AI words, Olivia Cal's 2026 AI-tells breakdown, Michael Kraabel's 200-phrase list, and an ongoing Reddit thread where people trade the tells they catch.