Yes, AI can write in your voice. Give it enough of your real writing and it will match your sentence length, vocabulary and rhythm closely. The bigger risk is that it invents plausible stories about you, like a customer call or a coffee conversation that never happened. To make AI write like you safely, feed it your real material, forbid made-up anecdotes, and read every draft as the person who lived it.
AI can write in your voice. That problem is mostly solved. Paste in a few dozen of your posts, or better, some transcripts of you talking, and a good model will pick up how you open, how long your sentences run and which words you'd never use.
The problem nobody warns you about is what comes next. The AI starts writing things that sound exactly like you and never happened. "A founder told me over coffee last week..." Beautiful sentence. There was no coffee. There was no founder.
I build an AI system that drafts posts for founders, and this is the failure I worry about most. It's worse than a clunky sentence, because it looks perfect. Here's why it happens and how I stop it.
Can AI really write in your voice?
On style, yes. On substance, only as far as you let it.
Voice has layers, and AI is uneven across them:
| Layer | What it means | Can AI do it? |
|---|---|---|
| Style | Sentence length, rhythm, vocabulary, how you open and close | Yes, reliably, from your past writing |
| Opinions | What you believe about your market and your work | Partly, if you've written or said it before |
| Stories | Things that actually happened to you | Only if you give it the story. Otherwise it invents one |
| Judgment | Why this post, this week, for these people | Not on its own |
Most tools sell the first row, because it demos well. We once ran an AI on a CEO's last 50 posts and matched his style almost perfectly. He read the drafts and said they sounded like every CEO on LinkedIn. The voice was his. The life inside it was missing.
The third row is where things go from generic to dangerous. When the AI doesn't have your story, it fills the gap with something plausible.
Why AI makes things up in your LinkedIn posts
Language models are built to produce the most likely next words. A LinkedIn post "should" have a story in the second paragraph, so the model writes one. It knows the shape of a founder anecdote (a meeting, a surprising line from a customer, a lesson) and it produces that shape with details that fit. It has no way to know the details are false, and it won't flag them.
When I was building our drafting system, I wrote this rule into the instructions almost word for word:
Don't try to predict something very specific about a story that happened to the client if you don't have any basis or factual information that this story actually happened. Use the first ideas to talk about principles, and about a point of view the client holds based on his previous posts. There is a high likelihood the AI is hallucinating.
I wrote it because of what happens on the client's side. The founder reads the draft and says "no, that's not my story," and they're disappointed. Worse than disappointed, actually. They stop trusting every other draft too, including the good ones.
And if a made-up story does slip through and gets posted, the people who catch it first are the ones closest to you. Your cofounder knows you didn't have that call. Your team knows the product never broke the way the post says it did. A CEO can't post about a conversation that never happened, because their own team will call it out.
The likeness gap: close enough is worse than wrong
I ran into the same failure with images before I understood it in text.
I tried generating pictures of myself for posts, using Gemini with reference photos. The results looked vaguely similar to me. Some looked very strange. None of them looked like me.
What struck me was that "vaguely similar" was worse than a picture of a stranger. A stranger is obviously a stock image. Someone who is almost you is unsettling, and people who know you notice immediately. We stopped using AI portraits and moved to text cards and concept images, formats where likeness doesn't matter.
Text works the same way. An invented story that's almost the kind of thing you'd say is worse than no story. Readers who don't know you may not catch it. The ones who do know you (your customers, your investors, your team) are exactly the audience you care about most.
There's a trust angle beyond your own network, too. LinkedIn's CEO, Ryan Roslansky, has said publicly that LinkedIn's built-in AI writing suggestions were less popular than he expected, and his explanation was that a LinkedIn profile is "your resume online." People are careful about what goes out under their name there, and they should be.
How to make AI write like you (without making things up)
Here's the setup I use and recommend. It works in ChatGPT, Claude or any decent tool.
- Build a voice sample from speech, plus your posts. Your posts show how you write when you're polishing. Transcripts of you talking show how you think. I get most of my raw material from voice memos, because it takes two minutes and captures real detail.
- Keep a story bank. A plain document of things that actually happened: the customer call, the hire that didn't work, the number that surprised you. One or two lines each, with a date. The AI may only use stories from this bank.
- Write the rule down explicitly. "Never invent anecdotes, conversations, customers, numbers or events. If a post needs a story and none is in the bank, write the principle instead and mark where a story should go." Put it in the system prompt or the instructions file, and keep it there.
- Ask for placeholders. A draft that says
[STORY: a time a customer pushed back on price]is more useful than a draft with a fake one. It tells you exactly what to add from memory. - Let the AI argue your point of view. Principles and opinions you've already stated are safe ground. "I think most founders hire sales too early" is something the AI can expand on without inventing facts.
- Read every draft as the person who lived it. For each specific detail, ask: did this happen, and could I tell the story at dinner? If not, cut it.
Step 6 is non-negotiable. No setup removes the need for a human gate. It just makes the gate faster.
If you want to check the style side separately, run a draft through an AI-sounding post checker. That catches the surface tells. It won't catch an invented story, which is why the story bank exists.
Principles first, stories from you
The rule that made our drafts better, and it surprised me, was letting the AI lead with principles.
When the AI doesn't have a real story, the first draft talks about a point of view: what you believe, why, what most people get wrong. Then you add the story, because you're the only one who has it. In practice this is a better division of labor than asking the AI for the whole post:
- The AI is good at: structure, a sharp first line, tightening, turning a messy thought into a clear argument, matching your cadence.
- You are the only source for: what happened, who said what, the real number, the stakes.
A draft built this way takes you a few minutes to finish, and everything in it is true. A draft with a fabricated story takes longer, because you first have to notice the fabrication, then delete it, then rewrite around the hole.
This is also why I think the most useful thing you can give an AI is raw material. Better prompts help far less. I wrote more about that in why AI posts sound like AI: readers spot the absence of anything only you could have written. Real stories are the cure for that, and you're the only supplier.
Signs your AI is putting words in your mouth
Watch for these in any draft, from any tool:
- Unnamed people with suspiciously specific quotes. "A CFO told me last month, 'We don't buy tools, we buy outcomes.'" Did one?
- Round, confident numbers you don't remember. "We cut onboarding time by 40%." Check before you post.
- Scenes with setting details. Coffee, airports, conference hallways. Models love a setting.
- Lessons that arrive too neatly. Real stories are messier than the AI's version.
- Experiences outside your life. A post about "when I was managing a 200-person team" when you've never managed 200 people.
Any one of these is a flag, and none of them means the draft is useless. Delete the invented part, keep the argument, add your real example.
Is it OK to let AI write in your voice at all?
I think so, with one condition: you stay the author. Using AI to draft is a tool choice, the same way a designer uses software. What matters is whether the post is true and whether it's yours.
The founders I've seen do this well share a habit. They spend their time on the input (capturing ideas, telling the story, deciding what they believe) and let the AI do the assembly. Then they read the output like it's going out under their name, because it is.
My whole view of leadership in the AI age comes down to being hands-on: you can't lead what you've never touched. That applies here. The AI can hold the pen. You have to supply the memories.