How do you write LinkedIn posts with AI that don't sound like AI? Stop fixing the words and start fixing the inputs. AI text reads as AI because of what it lacks: stakes, timelines, real events, real numbers. No word list solves that. The fix is grounding every draft in your actual work (a call, a voice note, a thing that happened Tuesday), drafting in your extracted voice, and keeping a human gate before anything ships. I call the underlying problem the specificity deficit, and it is the whole game.
Humanizer tools and "never say delve" lists polish the surface layer. Readers detect AI at a deeper layer: generic substance. A post with a real date, a real number, and something at risk reads as human even when AI drafted every sentence. Ground your drafts in your real work, let AI do the assembly, and judge the output, not the tool.
The voice match that failed perfectly
We ran our AI on a CEO's last 50 posts. It matched his voice with scary precision. Cadence, vocabulary, sentence structure: identical. I was proud of it.
Then he read the drafts and said: "This sounds like every CEO on LinkedIn."
He was right. And it took me a while to understand why, because by every metric I was optimizing, the drafts were perfect. The voice was his. The problem was that voice was never the problem. The drafts had his rhythm and nobody's life. No specific customer, no specific week, no specific mistake. They were built from his style plus generic knowledge, and generic knowledge is the same for every CEO on the platform. Match the voice on top of that and you get a beautifully personalized version of nothing.
The specificity deficit
Here is my claim: readers do not detect AI vocabulary. They detect the absence of anything only you could have written.
Think about the last post that made you stop. I would bet it had at least one of these: a number that was too specific to be invented, a timeline ("we shipped it March 3rd, it broke March 4th"), something the author risked or lost, an event that clearly happened. Now think about the posts you scroll past. They are not badly written. They are unattached. Five lessons that apply to everyone, which means they happened to no one.
That is the specificity deficit. AI models produce it by default because they draft from the average of everything they have read, and the average of everything contains no Tuesday. When you prompt "write a LinkedIn post about hiring lessons," the model has no choice but to hand you the consensus. It is not lying. It just has nothing at stake.
Which is why the entire humanizer industry is aimed at the wrong layer.
The wrong fix vs. the actual fix
Humanizer tools vary sentence length, swap flagged words, add contractions. Fine. But a generic post with varied sentence lengths is still a generic post. You cannot reword your way into specificity, because the specifics were never in the draft to begin with.
| The AI "tell" | The wrong fix (surface) | The actual fix (grounding) |
|---|---|---|
| Lessons that apply to everyone | Swap "delve" and "leverage" for plainer words | Draft from one real event: what happened, when, to whom |
| No numbers, or suspiciously round ones | Add hedges to sound casual | Use your real numbers, ugly ones included |
| No timeline, floating in "recently" | Vary sentence rhythm | Name the week. "Two months ago" beats "recently" every time |
| Nothing at stake for the author | Inject "personal" filler phrases | Say what it cost you: money, time, ego, a customer |
| Symmetrical listicles, every point equal weight | Reorder the bullets | Take a position. Kill the points you don't actually believe |
Surface fixes make AI text harder to flag. Grounding makes it worth reading. Only one of those builds an audience.
The market data backs this up. Originality.AI's 2025 study of 3,368 posts from 99 influential LinkedIn profiles classified 53.7% of long posts as likely AI-generated, and found that likely-AI posts underperform human-written ones in most professional sectors, by up to roughly 80% in strategy and innovation topics. The one category where AI won: Leadership & Inspiration. That is not a coincidence. It is the one genre where generic was already the norm, so the specificity deficit costs nothing. Everywhere readers expect substance, the deficit shows and the posts lose.
Stop shaming the tool. Judge the output.
I want to be precise about what I am not saying. I am not telling you to write by hand.
A developer who refuses an IDE is behind. A designer who refuses Figma is behind. And a writer who refuses AI, for brainstorming, for drafting, for grammar, is behind too. Using AI to write has nothing to do with authenticity. It is a professional tool, like every professional tool before it. Nobody asks a designer to hand-draw a SaaS product to prove they mean it.
The shame exists because the flood is real: people who never learned to say anything are now generating everything, and the output is landfill. But the right response is not to shame the tool. It is to judge the writing. Good content versus lazy content. That is the only distinction that ever mattered, and it survives every technology shift.
AI-assisted beats AI-generated
So here is the vocabulary I actually use, and the takeaway of this post.
AI-generated content starts from a prompt and generic knowledge. The machine supplies both the substance and the words. Result: every-CEO-on-LinkedIn syndrome, at scale.
AI-assisted content starts from you. My raw material is voice notes I record while walking, customer calls, things that actually broke this week. The AI's job is assembly: take the real event, draft it in my extracted voice, and hand it back for my one-tap yes or no. The substance is mine, and it could not have come from anyone else's model, because it came from my Tuesday.
Same technology. Opposite inputs. Opposite results. I have published this way to 4 million LinkedIn impressions in under a year, and I would guess most readers never once thought about whether AI touched the sentences, because the posts pass the only test that matters: only I could have written them.
The three-part loop, if you want to build it yourself:
- Ground. Every draft starts from a captured real event: voice note, call summary, incident. Never from a topic prompt.
- Voice. Draft in a voice extracted from your own past writing, not "professional LinkedIn tone".
- Gate. A human approves, tweaks, or kills every post. The gate is where the last 20% of judgment lives, and it is not optional.
The loop, pre-built
You can wire this with a transcription app and a good prompt library. If you want it pre-built, this loop is what Liftli is: it runs inside the AI you already use (Claude today, ChatGPT and Cursor next), drafts from your voice notes and calls in your extracted voice, and nothing ships without your tap. Free tier, no card.