How to Make AI Writing Sound Human

Readers on LinkedIn have learned to spot AI-written posts, and they usually do it before they reach the point. Nobody runs a detector for that. People notice the shape of the sentences, like a setup that promises a twist or a closing question nobody wants to answer.

In July 2026, Originality.ai analyzed 5,000 long-form LinkedIn posts and classified 81% of them as likely AI-generated (source). Take the number with care: the company sells AI detection, and detectors make mistakes. The direction, though, matches what you see in your own feed.

I noticed it on my own posts first. I drafted with ChatGPT, cleaned things up, and the result looked fine on screen. Then I read one out loud, and it didn't sound like me. Nobody talks like that. Every sentence was polished, every thought was sorted into a list with an emoji in front of it, and the whole thing read like a report. When you read a post, you want to hear a person, not a research paper.

This guide covers the patterns that give a post away, why the usual fix doesn't work, and what to do instead.

The patterns readers recognize

One pattern on its own proves nothing. People write "in conclusion" too, and some of the best writers on LinkedIn love the em dash. What gives a post away is density: four or five of these in 200 words, and the reader stops trusting the rest.

The list below comes from three places: the field guide that Wikipedia editors use to clean up AI text, academic studies of how instruction-tuned models write, and the drafts we have reviewed while building Elynd.

How the sentences are built

1. "It's not X, it's Y." The most common tell across every guide we've read. Example: "This isn't about productivity. It's about respect." The post denies something the reader never believed, then "reveals" the real point. To spot it, look for two sentences in a row where the first one negates and the second one asserts. An ordinary "not this but that" inside one sentence is fine.

2. "Not X. Not Y. Just Z." A cousin of the first one, built from fragments. Example: "No funding. No team. Just a laptop and a deadline."

3. Chopped fragments. Three or more sentences of one to three words each. Example: "Two years. No vacation. Alone." Used once in a long piece, a fragment adds weight. Used in every paragraph, it reads like a trailer voiceover.

4. A question answered immediately. Example: "The result? Zero signups." or "Why? Because nobody asked the customer." If the next sentence answers the question, you didn't need the question.

5. The teaser line. "Here's the thing:", "Here's the kicker", "And this is what I learned:", "Spoiler:". Each one announces that something interesting is coming instead of saying it.

6. The reassurance line. "And that's okay." It closes a paragraph about a struggle and lands almost word for word in thousands of posts.

7. The closing question to the audience. "What do you think?", "How do you handle this in your team?" A real question at the end is fine when the post earns it. A generic one is there because the model was told that engagement matters.

Which words get picked

8. Inflated significance. Pivotal, game-changer, a testament to, plays a crucial role. Wikipedia's guide lists this "importance inflation" as one of the most reliable signs (TechCrunch summary).

9. The trailing -ing clause. "...the team shipped on time, highlighting the importance of clear ownership." A study in PNAS by Reinhart and colleagues found that instruction-tuned models use present participial clauses two to five times as often as human writers (CMU news).

10. Filler openers. "It's worth noting", "Let's dive in", "In today's fast-paced world". They take up the first line of a paragraph and say nothing.

11. The announced conclusion. "In conclusion", "The takeaway:", "Bottom line:" at the start of a paragraph. A post is short enough that the reader can find the conclusion without a sign pointing to it.

How the post looks

12. Counted and decorated lists. "5 lessons I learned from...", a check mark or an arrow in front of every line. A list is fine when the content really is a list; the trouble starts when every thought in the post becomes a bullet.

13. Groups of three. "Fast, honest and drama-free." Humans use triads too; models reach for them by reflex, often with one word that adds nothing.

14. The em dash. It gets the most attention and is the weakest signal. Many good writers use it, and in Russian it's simply correct punctuation. It's also the easiest pattern to remove with one line in a prompt, which is why removing it changes little on its own.

15. Even rhythm. Every sentence about the same length, every paragraph one or two lines. Real writing speeds up and slows down. A long sentence that wanders a bit before it lands is often the most human line in the post.

Why an AI detector won't settle it

If you're tempted to paste your post into a detector and trust the percentage, don't. The Wikipedia editors who remove AI text from the encyclopedia at scale describe automated detectors as close to useless and rely on specific phrasing instead (same TechCrunch summary).

That's the approach in this guide too. Instead of a score, you need to see which lines read as a template, so you can decide what to do with each of them.

Why "write naturally" doesn't fix it

The usual advice is to add an instruction to the prompt: write in a natural, human tone, avoid clichés, sound authentic. We tested how models respond to instructions like these while tuning Elynd's own prompts, with 20 generated posts per variant. Three results changed how we write prompts.

The model copies examples. When a prompt contained a sample sentence in quotes ("for example, start with: ..."), the model opened with that exact construction in 18 of 20 posts and filled the rest with the topic. So if you add "for example, write like this" to fix a template, you plant a new one.

Bans on specific phrases work. One prompt explicitly banned a closing question to readers. With the ban, 1 post in 20 ended with a question. Without it, 20 out of 20 did.

Asking the model to judge its own text doesn't. We had a line in a prompt that said, roughly: if any LinkedIn author could have written this post, rewrite it. It made no visible difference.

The practical meaning: "sound human" and "avoid clichés" ask the model to evaluate its own writing, and it can't do that reliably. A concrete list of phrases it must not use is something it can follow.

The real cause: nothing specific to say

Banning phrases helps with the surface. The patterns themselves come from somewhere else.

In one of our tests we gave the model two kinds of input about the same author. One was a short episode: an event with a date, a number and a decision. The other was a generic credential, like "5 years in product management". The episode made it into the post in 10 out of 10 drafts. The credential appeared in only 5 out of 20, and when it did, it sat there as decoration while the rest of the post stayed general.

A model can only rearrange what you give it; it has no access to your story. Give it a topic, and it fills the empty space with form: the reversal, the fragments, the question at the end. Give it what happened on Tuesday, and there's less empty space to fill.

Elynd helps you write posts like this from your own material - try it free

What actually works

Start from an episode, not a topic. Before you ask for a draft, answer two questions in a few sentences. What exactly happened: who was involved, when, what number or detail you remember. And what you did differently because of it. Paste the answers first, then ask for the post. The draft will still need editing, but it will be about something.

Ban specific phrases, and don't give positive examples. Copy the patterns from this guide into your prompt as a list of things not to write: "Do not use 'It's not X, it's Y'. Do not end with a question to readers. Do not write three fragments in a row." Leave out any "write like this" sample, because the model will copy it.

Edit with the list, then read the post out loud. The list catches the patterns. Reading aloud catches most of the rest, such as a line you'd never say to a colleague or a paragraph that sounds like a slide. If you stumble over a sentence, rewrite it the way you'd say it.

A checklist for your next post

• Read it out loud. Would you say it this way to someone at work?
• Look for two sentences where the first denies something and the second reveals the "real" point.
• Look for three or more short fragments in a row, and for questions the next line answers.
• Check the last line. If it asks readers a question, does the post give them something to answer?
• Count the bullets and emoji. If every thought is a list item, turn some back into sentences.
• Find one detail only you could know: a number, a date, the name of a tool, something a client said. If there isn't one, adding it will do more than rephrasing.

The last point matters most. Most of the patterns above disappear when the post has something real in it, because a specific story doesn't need a dramatic setup to hold attention.