A founder reads a draft his agency sent over, gets two lines in, and types back four words: "this sounds like ChatGPT."
It's the most common reaction I hear to AI-written founder content, and it isn't really about ChatGPT. It's about a voice going missing. The post is competent. It's clean. It also could have been written for any founder in his category, which means it was written for none of them.
So the real question isn't whether AI can write in your voice. It's what has to be true for it to hold your voice instead of erasing it. That line is specific, and most setups land on the wrong side of it.
The problem is a missing spec, not a bad prompt
AI sounds like you when it has a spec for what "you" means, and sounds like nobody when it doesn't. That's the whole thing.
Hand a model a blank prompt and ask it to write a LinkedIn post as a B2B founder, and it does exactly what it's built to do: it fills the gap with its average. Average vocabulary, average structure, average opinion. The output isn't wrong. It's just the center of a very large distribution, and your voice was never the center. It was the edges.
People treat this as a prompting problem. Better instructions, a sharper system prompt, a few example posts pasted in. That helps at the margin. It doesn't fix the core issue, which is that the model has no documented account of how you specifically think and talk, and no raw material from you to draft from.
The fix is a voice profile: a written document of your phrasing, your recurring arguments, the words you reach for and the ones you'd never say, the positions you'll defend in public. Built from your recorded thinking, not guessed from three old posts. Once that exists, the model isn't inventing a founder. It's drafting against a real one.
The real danger: AI sands off the edges
The thing to worry about isn't robotic output. It's the opposite. It's AI making you sound smooth, balanced, and forgettable, by quietly removing the roughest, most specific parts of how you talk.
Your voice cuts through because of its edges. The contrarian position you hold that half your market disagrees with. The blunt phrasing you use that a committee would soften. The one example you keep coming back to because it actually happened to you. Those edges are what a reader remembers and what makes a post unmistakably yours.
A model optimizing for fluent, agreeable prose treats those edges as noise. It rounds the strong claim into a reasonable one. It swaps your specific story for a general principle. It adds a tidy, encouraging lesson to the end because that's what the average LinkedIn post does. Each individual edit looks like an improvement. The sum is a founder who sounds like a brand account.
This is why "it sounds like ChatGPT" is a precise complaint, not a vague one. The reader is detecting absence: the absence of the specific thing that made your voice worth reading.
Where AI actually belongs in the pipeline
AI belongs at the draft stage, working from your recorded raw material against your voice profile. It does not belong at the final post. That's the line.
Here's the order that holds the voice:
- You talk. A live extraction call captures how you actually think about a topic, in your words, including the tangents and the strong opinions. That recording is the raw material.
- AI drafts from that material. Not from a blank prompt. From your transcript, against your documented voice profile. The model compresses production work; it isn't sourcing the thinking.
- A human edits. Someone who knows your voice reads the draft and pulls it back toward you wherever the model drifted to its average.
- You approve. Nothing publishes until you've read it and said yes.
The order matters more than any single tool in it. AI in the middle, doing production, with your raw material on one side and human judgment plus your sign-off on the other. The failure mode in the market is AI on both ends: AI generating from nothing, and AI or nobody reviewing before it ships. That's the setup that produces the four-word reply.
The tells of AI-slop founder content
You can spot ungoverned AI founder content from the first two lines. The tells are consistent, and a voice profile plus a human edit removes each one.
| The tell | What's actually happening | What removes it |
|---|---|---|
| Words like delve, landscape, tapestry, harness | The model's house vocabulary, not yours | Voice profile lists your banned words; human catches strays |
| Perfectly even paragraphs, every post | Average structure, no human rhythm | A person rewrites for the way you actually talk |
| A neat lesson tacked onto every post | The model defaults to a takeaway | Editor cuts the forced moral; not every post has one |
| Opinions sanded to consensus | Fluency optimization rounds off the edge | Profile records the positions you'll defend |
| Generic examples, no real specifics | The model had no raw material from you | Drafts come from your recorded calls |
None of these get fixed by a better prompt. They get fixed by the model having your actual material to work from, and by a human reading the output who knows the difference between your voice and a competent imitation of it.
The human gate is the system, not a step
Cutting the human review is where AI founder content fails, every time. The gate isn't a polish pass you can drop to save money. It's the part of the system that does the work the model can't.
My practitioner read, not a measured figure: AI gets a draft most of the way to a founder's voice, and a human who knows that voice supplies the rest. The model handles speed, structure, and a first pass at tone. The person handles the line you'd actually say, the edge the model smoothed, the call on whether this even sounds like you. That last judgment is the whole job, and a model grading its own output can't make it.
This is also why "AI vs. human" is the wrong frame. It was never a choice between the two. It's AI doing the production a human shouldn't waste hours on, and a human doing the judgment AI can't do. Take the human out and you get volume with no voice. Take the AI out and you get voice with no cadence, which is why most founders quietly stop posting.
At DUO this runs on the Content Lab platform: the voice profile, the drafts, and the review gate live in one place, and nothing reaches your feed until you approve it.
The Upshot
If your AI-drafted founder content sounds generic, you don't need a better generator. You need two things almost nobody puts in place: a documented voice profile built from your own recorded thinking, and a human who knows your voice standing between the draft and the publish button.
The micro takeaway: judge any founder-content setup by where the AI sits. If it's generating from a blank prompt and shipping with no human edit, you'll get the four-word reply. If it's drafting from your recorded material against a written voice profile, with a person editing before you approve, it can sound like you, consistently, at a cadence you couldn't hold alone.
The sharper point: the flood of generic AI content didn't make a real founder voice less valuable. It made it the thing that stands out, because so little of what scrolls past has one. The edges a careless model sands off are the entire reason anyone reads founder content in the first place. Protecting them isn't a nice-to-have. It's the job.
For more on why ungoverned AI output drifts, see why your AI content sounds generic and what a human review gate catches. For the system this sits inside, read the AI content systems guide.
Common questions.
Can AI write LinkedIn posts that still sound like me as the founder?
Yes, but only with two things most setups skip. First, a documented voice profile: your phrasing, your recurring positions, the words you use and the ones you never would, written down. Second, a human review gate before you approve. With both, AI drafts from your own recorded thinking and a person catches the places it drifted. Without both, AI defaults to its average register and the output reads like everyone else's.
Why does AI-written founder content sound generic?
Because the model is filling gaps with its average, not your specifics. With no voice profile and no source material from you, it guesses tone from nothing and reaches for safe, balanced, conference-panel phrasing. The tells are consistent: words like delve and landscape, perfectly even paragraph lengths, a tidy lesson at the end of every post, opinions sanded down to consensus. A voice profile plus a human edit removes those.
What is a voice profile in founder content?
A written document of how a specific founder actually communicates: sentence rhythm, recurring arguments, vocabulary they use and vocabulary they avoid, the positions they'll defend, the topics they won't touch. It's built from recorded extraction calls, not guessed from a few old posts. The AI drafts against it. The human edits against it. It's the spec that makes the output reproducible instead of random.
Should a human still edit AI-drafted founder posts?
Always. The human gate is the system, not an optional polish step. AI gets a draft most of the way to your voice. A person who knows that voice supplies the rest: the line you'd actually say, the edge the model smoothed away, the take that's yours and not the model's average. Removing the human is where AI founder content fails.
