How to Write LinkedIn Posts That Don't Sound Like AI (A Ban List and a Method)

LinkedIn posts sound like AI mostly because of what they are about, and only partly because of how they are written. A post on “five lessons from leadership” is generic whether a person or a model typed it. The fix is to start from something that happened to you this week, say it out loud, and use AI only to tidy the structure. This article gives you the tell list, a voice-note method that takes about 25 minutes per post, four prompts to copy, and a before and after rewrite.

We have a stake in this. We write LinkedIn content for owners, and the question every sceptical owner asks is, “Isn’t this just AI slop? Will it sound like me?” The honest answer is that it sounds like you only if the raw material is yours. Our diagnosis: most of the posts doing well on LinkedIn right now look like each other, and it is the what, the topic, that gives them away.

Quick answer

How do you stop LinkedIn posts sounding like AI? Write about something specific that only you could say, record yourself telling it, and use AI for a rough draft only. Then edit by hand against a ban list: no “it’s not X, it’s Y” sentences, no runs of one-line paragraphs, no em dashes, no stock words, and no line another company could paste into its own post. Read it aloud. If you would not say it, cut it.

Why does so much LinkedIn content sound the same?

Because many people are using the same tools with the same prompts on the same topics. A 2024 analysis by Originality.ai, a vendor that sells AI detection, estimated that 54% of long LinkedIn posts (at least 100 words) were likely AI-written, using a sample of 8,795 posts. Detectors are imperfect and “likely AI” includes human posts that were heavily AI-edited, so treat the number as a direction. The direction matches what you see when you scroll.

There is also hard evidence on vocabulary. A 2025 study in Science Advances by Kobak and colleagues looked at more than 15 million biomedical abstracts and found sharp jumps in particular words after large language models arrived. “Delves” appeared about 28 times more often than expected, “underscores” about 14 times and “showcasing” about 11 times. The study is about scientific abstracts, not LinkedIn, but it shows that models have habits, and readers learn those habits quickly.

Readers do not need a detector. They have a nose. A simple version: if I see a post written the way everyone is writing them, I scroll straight past. That reaction is your real test, and it costs you nothing to run it.

What are the tells? The ban list

These are the patterns we remove on every draft. None of them proves a machine wrote the sentence, since people write this way too. Together they make a reader stop trusting you.

Same idea, two versions. The right-hand column is what we aim for.

Tell	What it looks like	Why it grates	Fix
Negative parallelism	“It’s not about X. It’s about Y.”	A fake contrast that sets up a point no one argued	State Y. If X is a real belief someone holds, name who holds it
Staccato runs	Six one-line paragraphs in a row	Reads like a slogan generator and wastes the fold	Vary sentence length. Join some lines. Keep one short line for emphasis
Em dashes	A dash in every second sentence	The most quoted tell among the people we work with	Use a comma, a colon or a full stop
Stock words	delve, leverage, unlock, landscape, seamless, game-changer, quiet, supercharge, navigate, tapestry	Model favourites, and nobody says them out loud	Say the plain version: use, open, market, smooth
Rule of three	“Clarity, confidence and consistency.”	Padded lists that sound complete and say little	One real point. If three are real, give each an example
Hollow openers	“In today’s fast-paced world…”	Throat-clearing before the first useful line	Start with the claim or the story
Vague authority	“Experts agree…”, “Research shows…”	No source, no weight	Name the study, or drop the line
Closing question bait	“Agree? Let me know below.”	Sounds like an engagement script	Ask a real question you do not know the answer to
Emoji bullets	A rocket for every point	Reads as a template	Use plain numbers

A word on the ban words. Our own house list also includes “quiet” and a few others because they appear in so much machine text, and we strip them on sight. Each word is fine in small doses. The trouble is how often they turn up.

The sneaky ones. When we tell a model to avoid “not X, but Y”, it often replaces it with “rather than X, Y” or “less X, more Y”. We tell the model to avoid the pattern, then we check by eye, because sometimes it is quite sneaky. If you see a sentence that sets up a fake opponent, rewrite it.

What is the method? Voice notes first, AI second, you last

Record yourself telling the story, let AI turn the transcript into a rough shape, then rewrite every line yourself. The order matters. If the model sees only a topic, it invents. If it sees your spoken words, it can only rearrange them.

Total time per post: about 25 minutes, most of it the hand edit.

Step 1: pick something that happened (2 minutes)

Take one real event from the last fortnight: a customer call, a mistake, a number that surprised you, a decision you made. Not a topic. “Why trust matters in sales” is a topic. “A buyer asked me to drop my price on a call and I did something I regret” is an event.

Mistake: starting from “what should I post about leadership?” The answer will be what everyone else posted.

Step 2: record a 2 to 3 minute voice note (5 minutes)

Use the voice recorder on your phone. Tell it to a friend: what happened, what you did, what you think about it now. Say the awkward bits. Do not tidy as you speak.

Raw beats polished here. We see content spoken into a mic outperform polished, make-up-ready posts, even on paid media. Your hesitations and half-sentences are the voice you are trying to protect.

Step 3: transcribe and get a rough draft (5 minutes)

Run the recording through any transcription tool, then paste the transcript into your AI model with this prompt:

Prompt 1: structure only “Here is a transcript of me telling a story. Turn it into a LinkedIn post of 180 to 250 words. Rules: use only the facts, numbers and phrases in the transcript. Add no new examples, no statistics and no advice I did not give. Keep my wording wherever possible. Do not use em dashes. Do not write any sentence of the form ‘it’s not X, it’s Y’ or ‘not just X but Y’. Do not write more than two one-line paragraphs in a row. Do not end with a question to the audience. Start with the most surprising specific detail.”

The aim is a first draft. Our own process is the same: a first draft with AI, a couple of prompts, then a heavy human edit.

Step 4: hand edit against the ban list (10 minutes)

Print it or read it aloud. For each sentence ask: would I say this to a customer? Replace anything that sounds like a brochure. Put back any phrase from the transcript that the model smoothed out. Run the ban list table above. If you want a second pair of eyes, use this:

Prompt 2: the tell checker “Here is my draft LinkedIn post. List every sentence that matches any of these patterns: negative parallelism (‘not X, it’s Y’), three or more short lines in a row, em dashes, stock words (delve, leverage, unlock, landscape, seamless, game-changer, quiet, navigate), rule-of-three lists, vague authority (‘experts say’), hollow openers. Quote each sentence and give a plain rewrite. Do not rewrite the whole post.”

This prompt produces a list. You decide what to change.

Prompt 3: hook options “Give me five alternative first two lines for this post. Each must contain a specific detail from the post (a number, a name of a situation, a surprise). None may state the lesson. None may start with a question.”

Prompt 4: the other model’s view “You are a busy buyer in [your industry] scrolling LinkedIn. Read this post. At which sentence would you stop reading, and why? Which sentence could have been written by any consultant? Quote them.”

A second model checking the first model’s work catches different habits. Treat its output as a hint.

Step 5: the scroll-past test (3 minutes)

Run the eight questions below. Any three yeses means rewrite.

Answer as a busy stranger, not as the author.

Can you see a before and after?

Here is a made-up example from a fictional manufacturer. The first version is the kind of post we see every week. The second is what it becomes after the method.

Before (the draft a model writes from a topic):

In today’s fast-paced business landscape, trust is everything.

It’s not about price. It’s about relationships.

Last week I had a conversation with a customer.

It changed my perspective.

Here are three lessons I learned:

1. Listen first
2. Be transparent
3. Deliver consistently

Because at the end of the day, great service isn’t just a transaction. It’s a partnership.

Agree? Let me know in the comments!

It has the hollow opener, the stock word, the fake contrast twice, staccato lines, the rule of three, the closing bait, and no fact anyone could check. (Add an em dash somewhere and the set is complete.) It could be pasted into any firm’s feed.

After (from a voice note about what happened):

A buyer rang us last Tuesday and asked us to drop a quote by 8%. I said yes before he had finished the sentence.

It took me an hour to work out why that bothered me. We had spent three weeks on that quote. He had never once asked what was in it. We had also never told him.

So I rang him back and walked him through the three lines that made up most of the price: the heat treatment, the inspection report he needed for his own customer, and the six day lead time. I did not offer another discount. I asked him which of the three he could live without.

He kept all three. We kept the original price.

My mistake was letting a quote go out for three weeks without ever explaining it. We now put a one-page note of “what you are paying for” with every quote over a certain size.

What do you put in front of a buyer before they ask for a discount?

Same lesson, trust and transparency. The second version has a number, a day, a decision, a regret, a fix, and a closing question to which the author does not know the answer. Nothing in it could belong to another business. That is the whole trick: specific facts only the author has.

What does this look like as a weekly routine?

Two voice notes a week produce four or five posts, because one story often gives you a main post and a shorter follow-up from a different angle (written to stand alone, never “yesterday I wrote about…”).

A 90 minute weekly routine:

• Monday, 15 min: record two voice notes of 2 to 3 minutes each on events from the last fortnight.
• Monday, 15 min: transcribe and run Prompt 1 for each.
• Tuesday, 40 min: hand edit four posts, run Prompt 2, rewrite the hooks.
• Tuesday, 20 min: scroll-past test, schedule, bank.

If you cannot find the time, have someone interview you instead. Our LinkedIn personal branding playbook has the 12 question interview that gives you weeks of stories in one hour.

Longer often feels more human. In our experience, longer posts feel less machine-made, because a story has room for the odd detail. Test one longer post a week for a month and compare.

Be open about your process. Using AI for a draft is fine. Pretending you did not is the risk. If a reader asks, say how you work: “I talk, a tool tidies the draft, I rewrite it.” People forgive the tool. They do not forgive a post that was never yours.

Common mistakes

• Asking the model for ideas. It will give you the most common ideas, by definition. Ask it only to structure what you already said.
• Editing out the rough bits. A clumsy sentence in your own voice beats a polished one in everyone’s.
• Banning words but keeping the pattern. Swap “leverage” for “use” and keep the fake contrast, and the post still reads as machine-made.
• Posting a “lesson” with no story. A list of lessons without the event that taught them is a template.
• Posting daily to hit a streak. A weaker post a day does more harm than a stronger one on three days.

Questions people ask

Can you tell if a LinkedIn post was written by AI?

Often, though not reliably. Detectors make mistakes and human writers use the same patterns. Readers rely on feel: generic subject, fake contrasts, one-line paragraphs and stock words.

Is it okay to use AI to write LinkedIn posts?

Yes, as a drafting aid for your own material. Use your voice note as the source, rewrite by hand and check the claims. A person should own every post.

What words give away AI writing?

Words like delve, leverage, unlock, landscape, seamless and underscore show up more than they should in machine text. In a 2025 study of biomedical abstracts, “delves” appeared about 28 times more often than expected. The overused list shifts with each model generation, so use your ear as well as a list.

Does LinkedIn penalise AI-generated posts?

We have not seen LinkedIn publish a rule penalising them, and we cannot verify that its feed treats them differently. Our concern is simpler: generic posts get scrolled past.

Why do AI posts use so many short lines?

Short lines look punchy and are cheap to produce, so models repeat them. Run more than two in a row and the post starts to sound like a slogan.

Sources

• Originality.ai, “Over Half of Long Posts on LinkedIn are Likely AI-Generated Since ChatGPT Launched”, 2024 (vendor detector estimate; 8,795 posts of at least 100 words; “likely AI” includes heavily AI-edited human text).
• Kobak, González-Márquez, Horvát and Lause, “Delving into LLM-assisted writing in biomedical publications through excess vocabulary”, Science Advances, 2025 (excess word frequencies in more than 15 million PubMed abstracts).
• Wikipedia contributors, “Signs of AI writing”, accessed 2026 (observations of common patterns, with the caveat that humans write this way too).

Want us to do this for you?

If you would rather talk for an hour than write for ten, we interview you, draft in your voice and send every post for your approval before it goes live. Tell us about your business and we will show you a sample.