How to make AI LinkedIn posts sound human

AI LinkedIn posts sound like AI because they contain nothing that is true only of you. The model writes competent sentences about a topic anyone could write about, and readers detect that instantly, long before they notice any particular word. The fix is not a better prompt or a longer list of banned phrases. It is feeding the model raw material it could not have invented: a scene with a date on it, a number from your own business, a sentence a client actually said, an opinion you would defend in an argument. Everything else in this article is downstream of that. This matters more than it used to: an Originality.ai analysis of 5,000 long-form LinkedIn posts from July 2026 classified 81.2% as likely AI-generated, which means generic machine prose is now the background noise you are writing against.

Why do AI-written LinkedIn posts sound generic?

A language model produces the most probable continuation of what it has been given. Ask it for a LinkedIn post about hiring and it will produce the average of every LinkedIn post about hiring, which is a genuinely accurate summary of the topic and completely worthless as a piece of writing. It is not a failure of capability. It is a correct answer to a question nobody should have asked.

Four things go wrong in sequence, and they compound.

1. No specific stories

The model has no access to the Tuesday you lost a deal because you sent the proposal before the discovery call was finished. So it writes about the importance of discovery calls. Every reader has read that post two hundred times. The specific version has never been read by anyone.

2. Hedged claims

Models are trained toward balance and away from confident assertions, so an AI draft says “can often be an effective approach for many teams” where a person would say “this is wrong and here is what I do instead.” Hedging is what makes a post unarguable, and unarguable posts do not get comments. In our cohort the median top-decile post drew 72 comments against 5 in the bottom half, a far wider gap than reactions (235 against 80). Comments are the scarce signal and they come from disagreement, recognition, or a question worth answering.

3. Uniform rhythm

Human writing is lumpy. A twenty-four word sentence, then four words. AI drafts settle into a narrow band, usually twelve to eighteen words per sentence, paragraph after paragraph. Nobody consciously notices this. Everybody feels it. It is the single most reliable tell and the easiest to fix, because you fix it by deleting.

4. Giveaway vocabulary and shapes

Individual words matter less than people think, but a few have become so overused in machine output that they now function as a signature. More damaging are the structural habits: the three-item list where a real thought would be lopsided, the opening line that restates the prompt, the rhetorical question addressed to nobody in the last line, and the “it's not just X, it's Y” construction that appears in roughly every second AI-drafted post on the platform.

What are the exact tells readers notice?

Not a stylistic wishlist. This is the checklist we run drafts against, with the reason each item is a tell and the fix.

The vocabulary row is the one people fixate on, and it is the least important. You can strip every flagged word from a post that has nothing in it and it will still read as machine output, because the emptiness is the tell. Fix the input first, then run the word list.

What does the engagement data say reads as human?

We ranked 12,988 English-language LinkedIn posts from 65 creators by engagement rate and compared the top decile against the bottom half. The patterns line up almost exactly with the tells above.

Read the first row and the last row together, because they are the whole argument. First person nearly doubles at the top. Numbers do not move at all (32.1 against 32.8). A model can put a number in a hook without being told anything about you, which is precisely why numbers do not separate the groups. It cannot write a first-person sentence about something that happened to you without being told what happened. That is the difference the data is actually measuring: not a stylistic preference for the word “I,” but the presence of material only the author has.

Second person is the other half. Nearly half of top-decile hooks address the reader directly somewhere in the visible text, against a third of the bottom half. Specific about you, directed at them. Generic AI output manages neither: it writes about a topic, to nobody.

Here is what that looks like in practice. Both of these are real hooks from the dataset, pulled at the point LinkedIn cuts to “see more.”

What does a before and after rewrite look like?

Below is an illustrative example, not a post from the dataset. We wrote both versions to demonstrate the edit. The scenario is a fractional CFO writing about pricing.

Before: what a model produces from a topic alone

Pricing is one of the most important decisions a business can make. Many companies underprice their services without realizing the long-term impact on their growth.

Here are three things to consider when setting your prices: understand your true costs, research your market positioning, and communicate your value clearly to prospects.

Getting pricing right can often be the difference between a business that struggles and one that scales sustainably. It's not just about the number, it's about the confidence behind it.

What has your experience been with pricing? Let me know in the comments.

Nothing in that is wrong. That is the problem. It contains zero facts, three hedges (“many,” “can often be,” “sustainably”), one rule of three, one X-not-Y construction, and a closing question addressed to the general public. All eleven sentences sit between eleven and nineteen words.

The material the author actually had

Before the rewrite, the author spent four minutes answering three questions out loud into a voice memo:

• When did you last change a client's prices, and what happened?
• What did the client say when you proposed it?
• What did you believe about pricing five years ago that you now think is wrong?

Answers: a design agency, March, moved their project minimum from $12,000 to $28,000. The founder said “we will lose everyone.” They lost four of eleven active conversations and closed three of the remaining seven, which was more revenue than the previous quarter. And five years ago the author thought pricing was a spreadsheet exercise rather than a positioning one.

After: the same post with that material in it

In March I told a design agency to raise their project minimum from $12,000 to $28,000.

The founder said: “we will lose everyone.”

They lost four of eleven live conversations. They closed three of the remaining seven. That quarter beat the previous one on revenue with fewer than half the proposals written.

The part nobody tells you: the four who left were the four who had been asking for discounts since January. Raising the price did not lose them. It just told the truth faster.

I used to treat pricing as a spreadsheet exercise. Cost, margin, done. It is a positioning exercise, and the spreadsheet only tells you the floor.

If you are quoting the same number you quoted two years ago, you are not being competitive. You are being out of date.

Same topic, same length, same author. The edit added five facts (March, $12,000, $28,000, four of eleven, three of seven), one quoted sentence, one admission of having been wrong, and one claim the reader can disagree with. Sentence lengths now run from six words to twenty-eight. Not a single sentence was made more eloquent. This is the entire technique.

Note also what happened to the opening. The generic version opens by restating the topic. The rewrite opens mid-scene with a date and a number, which is the pattern our guide to opening a LinkedIn post covers in more depth, and which our hook analysis found separating the top decile.

How do you feed a model your real material?

The bottleneck is capture, not prompting. Almost everyone who complains their AI posts sound robotic is giving the model a topic and expecting a memoir. Four capture habits, in order of how much they return.

1. The ten-minute weekly voice memo. Immediately after a client call, answer three fixed questions out loud: what surprised me, what did I have to explain twice, what did I get wrong. Transcribe it. That transcript is worth more than any prompt library.
2. A running list of things people asked you.Real questions, in the asker's words. Each one is a post, and it arrives pre-validated because somebody actually wanted the answer.
3. Numbers you are allowed to publish. Keep a short file. Deal sizes, cycle lengths, headcounts, conversion rates, anything with a digit in it that will not get you in trouble. Vague posts are usually a permissions problem, not a writing problem.
4. Opinions with a named opponent. Write down the thing you believe that a respected peer would argue with. Feed the model the belief and the counterargument, and ask it to write the version where you take the position, not the balanced one.

Then structure the prompt around the material rather than the topic. A workable shape: here is what happened, in my own messy words. Here is the one line I want a reader to remember. Here are three words I never use. Write it in short paragraphs, open on the concrete detail, do not add anything I did not tell you. That last clause is doing most of the work.

Two prompt instructions that measurably help

• “Do not add any fact I did not give you.” Models fill gaps by default. Explicitly closing that door produces shorter, truer drafts and makes it obvious where your material ran out.
• “Vary sentence length aggressively. At least two sentences under six words.” Rhythm is one of the few tells you can fix by instruction rather than by editing.

How do you build a voice guide a model can actually use?

Most voice guides are useless because they describe a personality rather than a set of rules. “Warm but authoritative, approachable yet expert” gives a model nothing to act on. A usable voice guide is short, mechanical, and made mostly of prohibitions.

• Twenty words you use and twenty you never use.Not synonyms. Real preferences. If you say “buyers” and never “prospects,” “team” and never “folks,” write both columns down. This single page moves output quality more than any other instruction.
• Three sentences you would never write.Concrete examples of the wrong register beat any adjective. “I would never write: excited to announce.”
• Your default paragraph length. One line or four. Pick one and say it.
• Punctuation rules. Whether you use exclamation marks, whether you use bullet points inside posts, whether you use emoji at all. Be literal.
• Two of your own posts that worked, annotated. Not just pasted. Say which line did the work and why. A model given a labelled example generalises far better than one given ten unlabelled ones.

Keep it under a page. Voice guides fail by growing: past about 500 words, later instructions start diluting earlier ones and the output drifts back toward the mean. Revise it every few weeks by adding one rule from whatever you found yourself editing most.

Do emoji and hashtags make a post read as human or as automation?

Opposite answers, which surprises people. In our cohort, 24.3% of top-decile hooks contained at least one emoji against 2.7% of the bottom half. That is the widest single gap in the entire hook dataset, wider than first person, wider than question marks. We would not tell anyone to add emoji to raise engagement, because the causal story runs the other way: emoji travel with a conversational, personal register, and that register is what performs. The emoji is a symptom.

Hashtags go the other way. Four or more hashtags appeared in 25.9% of top-decile posts and 41.9% of the bottom half, about 1.6x more common at the bottom. A wall of tags is what people add instead of writing a good opening, and it is also the most visible marker of automated output, since generic tools append them by default. Our full analysis of hashtag use covers why the feature they were built for no longer really exists.

The practical rule that falls out of both: let the post be as informal as you actually are, and strip the mechanical decoration. Both moves point the same direction, which is away from output that looks assembled.

Where should AI never touch the post?

Four places. These are not stylistic preferences, they are the spots where machine output is actively costly.

• The first two lines. This is the only part most people will read, and it is the part a model is worst at, because a good opening depends on knowing which detail is surprising. Write these yourself even if a tool drafts everything else.
• Any number. Models produce plausible figures. Plausible is the exact failure mode you cannot afford in public, in your own name, about your own business.
• Anything about a named person. If a client, colleague or competitor appears by name, every word about them is yours. No exceptions.
• Opinions you have not formed. A borrowed opinion holds until the first person replies asking you to defend it. Then you are arguing a position you do not hold in front of your buyers.

Where AI genuinely helps

• Cutting. Models are far better at removing 40% of a draft than at writing one. Give it your rambling version and ask for it shorter with nothing added.
• Twelve hook options. Not to use one verbatim, but to see the shape of the space and notice which angle you had not considered.
• Structure.“Here is a story, where does it start?” is a question a model answers well.
• Consistency. The honest one. The largest effect any tool has on your results is that posts get written on the days you did not feel like writing. Talent that posts eleven times a year loses to competence that posts weekly.
• Format decisions. Native video over-indexed about 2.6x in our top decile (26.3% against 10.1%), while shared link posts over-indexed about 1.5x in the bottom half. A tool that reminds you of that at the right moment is worth something, and our data on how link posts perform gets into the contested details.

Does it matter that most LinkedIn posts now look AI-generated?

It matters, but not in the way the panic suggests. The Originality.ai figure above, 81.2% of 5,000 long-form July 2026 posts classified as likely AI, should be read with real caution: detectors produce false positives, they penalise plain and formal writing, and the sample was drawn from search results rather than the whole platform. We would not present that number as the true share of AI content on LinkedIn.

What it does establish is a direction. There is a great deal of undifferentiated machine prose in the feed, and its floor keeps rising as models improve. Two consequences follow. First, competent generic writing is now worth nothing, because it is free and infinite. Second, the things a model cannot supply have appreciated: your specific numbers, your mistakes, your unpopular positions, and the fact that you were in the room.

On the rules, for completeness: we could not find any LinkedIn policy prohibiting AI-assisted writing or requiring disclosure of it. The published policies target inauthentic identity and inauthentic engagement, not drafting tools. We went through the relevant clauses in our breakdown of what LinkedIn's terms say about automation. That is an observation about the documents, not legal advice.

A checklist to run before you post

1. Does this contain at least one fact that is true only of me? If not, do not post it.
2. Read the first two lines alone. Would you keep reading?
3. Count sentence lengths. Are at least two under six words?
4. Find every hedge (“can often,” “many,” “may help”). Delete or commit.
5. Is there one sentence a smart peer would argue with? If not, you will get reactions and no comments.
6. Run your banned word list. Then read it aloud, which catches the rest.
7. Would someone who has met you recognise this as yours? That is the only test that has ever mattered.

The short version of how to make AI LinkedIn posts sound human

Stop trying to make the prose sound human and start making the content be yours. A model given a topic writes the average post about that topic, which is exactly what 81% of the feed already looks like. A model given a March meeting, a $28,000 number, and a sentence a founder actually said writes something nobody else could have written, and then your only job is to fix the rhythm, cut the hedges, and write the first two lines yourself. The data agrees: first-person openers appear about twice as often at the top of our cohort, and the thing that makes first person possible is having been there. AI cannot supply that. It can do almost everything else.