Why do AI LinkedIn posts sound so cringe?

Because a language model is built to return the most probable next words, and the most probable phrasing is, by definition, the average phrasing. Average writing does not sound like a specific person. It sounds like the middle of everyone, which is exactly what makes it land as hollow.

This is worth sitting with, because it explains why the problem does not go away when you switch tools. Every general-purpose model is solving the same optimisation, so every general-purpose model drifts toward the same centre. Two founders in different industries, prompting different tools on different days, end up publishing posts that could swap bylines without anyone noticing.

The uncomfortable part is that the post is not badly written. It is competently written and belongs to nobody.

Readers pick this up faster than they can explain it. They will not think about probability distributions. They will feel a small drop in trust, scroll past, and never be able to tell you why. That reaction is the entire cost, and it is why the word people reach for is cringe rather than wrong.

Not the vocabulary. It is the absence of anything only you could have written. A generated post typically carries no numbers, no names, no dates, no cost, and no moment where something was genuinely at risk. Strip those out of a human post and it reads as generated too.

The most quoted giveaway is punctuation, the em dash in particular, and it is real enough that people now scan for it. But treating that as the problem is a mistake. Surface tics are the symptom. You can clean every one of them out of a post and still be left with something that says nothing, because the emptiness was never in the punctuation.

The deeper pattern is tidiness. Generated posts tend to arrive suspiciously well balanced: three points of roughly equal weight, each with roughly equal support, resolving into a conclusion nobody could argue with. Real thinking is lumpier than that. One point matters far more than the others, a caveat sits awkwardly in the middle, and the ending is often less tidy than the opening promised.

• It could sit under anyone's name in your industry without a single edit.
• Nothing in it could be wrong, because nothing in it is a claim.
• It advises rather than recounts. Nobody actually did anything.
• It resolves too neatly, with every loose end tucked in.

Read your last generated draft against that list. If it clears all four, the draft is not salvageable by editing, because there is nothing underneath the words to edit toward.

Almost always the prompt. Ask for a LinkedIn post about leadership and you have supplied no information at all, so the model has nothing to work from except the average of everything ever written about leadership. It returns exactly that, and it is right to.

The gap is that most people prompt with a topic when the thing that makes writing readable is a specific. A topic is leadership. A specific is that you promoted your best engineer into management, watched them struggle for five months, and moved them back with their pay unchanged. The first has one obvious post in it that has been written ten thousand times. The second has only ever been written by you.

This is also why longer prompts do not fix it. Adding be engaging and use a strong hook and avoid corporate language is still instruction about form, not information about you. You cannot request specificity into existence. It has to be supplied.

Yes, though not for the reason most people assume. There is no reliable evidence that LinkedIn detects and suppresses AI text as such. What it does measure is whether people stop, read, and respond, and generic posts fail that test on their own merits.

The mechanism is ordinary. A post that reads as filler gets a fractionally lower dwell time and fewer meaningful comments in its first hour, the distribution narrows accordingly, and it quietly dies. No penalty was applied. The post simply did not earn anything. We went through what the platform does and does not appear to act on in the piece on AI content and reach.

The compounding cost is worse than any single post. Publish generic writing for three months and you have taught your network that your posts are safe to scroll past. That habit is considerably harder to reverse than it was to create.

Supply the things a model cannot invent, then let it handle the assembly. Your numbers, your names, the objection you actually got, the thing that went wrong and what it cost. Specificity is the whole game, and it is the one input that has to come from you.

1. 1.Start from an event, not a topic. Something that happened, with a date attached.
2. 2.Put one real number in. Revenue, headcount, days, the price you paid, the percentage you were wrong by.
3. 3.Name the thing you believed before, and what changed it. Belief plus reversal is the most readable shape there is.
4. 4.Keep the sentence that sounds slightly too blunt. That is usually the only line that sounds like a person.
5. 5.Cut the closing lesson. Readers resent being told what to conclude, and the lesson is the most generic paragraph in almost every draft.

Do that and the writing has something to be about, which is the part no amount of prompting or editing can supply after the fact. It also gets faster with practice, because you stop hunting for topics and start noticing material in your own week. The five-pass edit that does most of the work is laid out in how to make AI writing sound human.

Then keep the assistance and change what it starts from. The trap was never using AI to write. It was starting from a blank prompt, which can only ever return the average. A draft built from your own material and your own way of putting things does not have that problem.

That distinction is the whole reason innernote exists. It learns how you actually write, then drafts from your material in that voice, so what comes back reads like a fast version of you rather than a competent version of nobody. You edit it, which takes a minute, instead of rewriting it, which takes twenty and usually ends with you closing the tab. There is a walkthrough of the flow on the product page.

If you want to test the difference on something you have already written, paste a recent draft into the AI text cleaner and see how much of it survives. Then start free and write next week's posts from your own week instead of from a topic.