LinkedIn Has an AI Problem. We Built an AI Tool to Solve It.

A user told us our tool sounded like AI. They were right. We checked their writing samples and every single one had been written by AI. “Navigating the evolving landscape.” “I’m excited to share.” The classic closer: “What’s your experience with this?”

The system learned their voice exactly as instructed. The voice it learned just wasn’t theirs.

Researchers call this model collapse. Train AI on AI-generated data and the outputs degrade — technically coherent but increasingly hollow. Our user had been using AI to write LinkedIn posts for two years. Those posts became their writing samples. We learned the average of what AI produces when writing about their topics. Then produced more of the same.

AI content trained on AI content trained on AI content. Garbage in, garbage out — in a loop subtle enough that nobody noticed.

Over 50% of LinkedIn long-form posts are now AI-generated

Originality.ai analyzed 99 influential LinkedIn profiles. Over 50% of long-form posts in 2025 were likely AI-generated. That number keeps climbing. In some industries it’s way higher. And human-written posts consistently outperform them on engagement. People can feel the difference between a post that sounds like a person and a post that sounds like “content,” even if they can’t explain why.

Which means for a huge chunk of professionals active on LinkedIn the last two years, their post history isn’t their voice anymore. It’s what AI sounds like when writing about their topics in a generically professional register.

And voice-matching tools (including ours, being honest here) that learn from post history are, to varying degrees, learning the wrong thing.

The industry has an uncomfortable recursion problem. Tools designed to help people sound like themselves are being trained on content that was designed to sound like content.

The three patterns you’ve absorbed without noticing

If you’ve been on LinkedIn for more than two years, you’ve internalized these whether you wanted to or not.

The narrative opener. “Three years ago, I made a mistake.” “I used to believe X. I was wrong.” This dominated feeds from 2021 to 2024. Every content coach taught it, every AI tool learned it, and now it reads as completely generic even when a real human writes it from the heart.

The listicle pivot. A story that builds to a neat numbered list. “Here’s what I learned:” followed by three tidy points. Clean. Structured. Completely emptied of anything that sounds like a specific person actually wrote it.

The engagement closer. “What’s your experience with this?” “I’d love to hear your thoughts.” These worked for about five minutes. Then they became noise. LinkedIn’s algorithm actually started penalizing them. And yet.. humans and AI alike still reach for them every single time.

These patterns spread like an accent. Research tracking word frequency found that even people who never touch AI tools have started imitating the AI register. Using phrases like “navigate,” “transformative,” and “it’s worth noting” because they’ve absorbed them from the feed.

So when someone reads their AI-generated post and thinks “yeah, that sounds about right”.. they’re not wrong. It does sound right. Right according to a standard that’s been quietly reshaped by two years of AI content everywhere.

Their calibration is off. And they don’t know it.

AI polishing makes it worse

When you use an AI tool to “clean up” a draft you already wrote using an absorbed template.. you’re running the same averaging function twice. You’re asking a model trained on generic LinkedIn to polish something you wrote in the style of generic LinkedIn.

The output comes out even more beige than either input. Your rough draft at least had some of your weird specific friction in it. The AI sanded it all off.

Your voice didn’t go anywhere

It just stopped being exercised publicly.

The way you write an email to a close colleague. How you explain something complicated over the phone. The story you tell at dinner about a frustrating day at work. How you text. How you talk.

That’s still your voice. Still in there. You use it every day in contexts where you’re not performing. It’s just not on LinkedIn anymore.

What’s on LinkedIn is the performed version. The one that got smoothed by AI, had its rough edges sanded down, its sentence rhythms replaced with statistically average ones. Closed with a question you would never actually ask out loud.

The voice isn’t gone. It’s just been locked out of the one place where it would matter most.

Speech is the way back

The worst place to find someone’s real voice is their recent LinkedIn posts. The best place? How they talk.

Speaking in conversation is fundamentally different from writing. You can’t edit before you send. You use the vocabulary you actually reach for. The hedges, the restarts, the qualifications. All involuntary expressions of how you relate to the idea you’re describing.

A twenty-minute interview transcript contains more authentic voice signal than two years of AI-assisted LinkedIn posts. Not because it’s more polished. Because it’s less.

Forensic linguist James Pennebaker spent decades showing that the most identifying features of someone’s voice are the involuntary patterns. Function words, hedging, sentence rhythm, pronoun use. Below conscious attention. Almost impossible to fake. And they survive in speech much better than in writing that’s been edited and cleaned up.

When someone does an interview with Outerview, we’re not asking them to remember what they sound like. We’re creating conditions where they can’t help sounding like themselves.

---

Outerview generates LinkedIn posts from voice interviews, not written prompts. Every post starts with a conversation. Start your first interview at outerview.app