How to train AI to write in your voice (not everyone else's)
Generic AI writing is recognisable because it was trained on generic content. Here's the practical process for giving a model enough of your real writing to finally sound like you.
The most common complaint about AI writing tools is that they all sound the same. And they do — because they were all trained on the same internet. The fix isn't a better prompt. It's giving the model a different dataset: yours.
This sounds complicated. It isn't. Here's what actually works.
What "training on your voice" actually means
You're not fine-tuning a model from scratch. What you're doing is giving the AI a set of real examples to pattern-match against, then scoring the output against those same examples to see how close it got. This is sometimes called retrieval-augmented generation, or RAG — but that term obscures what's actually happening.
Think of it like hiring a ghostwriter. The first thing any good ghostwriter does is read everything you've ever published. Not to copy it — to absorb your rhythm, your instincts, the way you build an argument. They learn which words you'd never use, which analogies you'd reach for, whether you tend to open with data or with a story.
AI can do the same thing, but it needs material. The more specific and real the material, the better the output.
What to give it
The best source is your own published writing — posts, threads, essays, emails you've sent that you're proud of. Three to five examples is enough to extract a style fingerprint. Ten is better. Transcripts of you speaking (podcast episodes, recorded talks) work too, especially if you write the way you talk.
What doesn't work: your website's marketing copy, posts you paid a ghostwriter to write, anything that was AI-assisted already. Those inherit other people's patterns. You want the most raw, least-edited version of your thinking available.
The four dimensions that define your voice
When a tool like Verbatrum reads your posts, it's specifically looking for:
- Hook pattern. Do you open with a question, a claim, a story, or a data point? Most people are consistent without realising it.
- Sentence rhythm. Short punchy sentences vs. long flowing ones. How you use sentence fragments. Whether you write in parallel structure or deliberately break it.
- Analogy domain. Where do your comparisons come from? Sports, cooking, physics, history? This is the most distinctive signal and the hardest for AI to fake without examples.
- What you never say. Negative signals are as important as positive ones. "Passionate about," "let's dive in," and "it's time to" are tells that this isn't you, even if everything else is right.
How to tell if it worked
Read the output out loud. If you'd feel comfortable saying it in a meeting with someone who knows your work, the voice match is good. If a sentence sounds like it belongs in a press release, the model defaulted to generic — it didn't have enough of your real writing to pattern-match against that sentence.
Verbatrum shows this as a Voice Match Score: a number from 0–100 measuring the structural distance between the generated post and your provided examples. When it's above 80, the post is usually ready. Below 60 means the generation didn't have enough signal — add more examples and run it again.
The practical setup
You only need to do this once. Paste five or more of your real LinkedIn posts (or other writing), and Verbatrum builds a persistent voice profile. Every tool in the suite then uses that profile automatically — so the second time you generate a post, you don't go through the setup again. The voice is already there.
The one caveat: if your writing style evolves significantly, update your examples. Voice profiles aren't set-and-forget forever — they're a snapshot of how you wrote when you created them.