Getting your AI to sound like you

I’ve fought the use of AI for writing for the longest time. Most of the time the thinking was right. The structure was right. The em-dashes were correct in the sense that they connected the right clauses. But reading it aloud, something landed wrong.

It fell just short of sounding like me. Even when I spent the time to ask it to “act as a leadership development professional.”

There was a slightly elevated, slightly performative, mildly aspirational English that had taken over professional writing over the last decade. Here’s the thing. What I’ve learned is… In today’s AI landscape… I realize the fatigue of that phrasing. The audience is tired of it; my own ear is tired of it. The AI, trained on the volume of writing that’s out there, defaults to it. And I’m the type of person who speaks corporate unironically in my day-to-day life.

And — I write differently than that. Not better — differently. My sentences run longer. I use specific verbs over generic ones. I don’t tell readers what I’m about to say; I just say it. When the AI hands me back a draft that smooths all of that out, what comes back isn’t me.

The advice I got from the internet wasn’t totally bad. Write in my voice. I’m a senior practitioner who values restraint and specificity. But — the AI hears “senior practitioner” and serves the average of senior-practitioner writing on the internet. And you end up with unspecified specificity — because AI uses prediction and pattern to pick the next word.

Then, a metaphor I heard changed everything for me. Treat your AI as your junior employee. I needed to let the AI learn my voice the same way a young writer learns a senior writer’s voice — by reading the work. Not the rules. The work. Give it words to help it predict.

Once a month I take thirty minutes and feed my AI assistant a fresh batch of my recent writing. Not edited. Not curated. The actual paragraphs from recent proposals, articles, emails to clients. I tell it: this is what I sound like right now. The next thing you draft for me should match this register.

The AI got specific information from this that no prompt could carry. It learned that I tend to use em-dashes even though my “make it sound not like AI rule” would’ve taken those out. That I almost never write passion or leverage or unlock with certain audiences. That when I’m explaining something complex, I usually start with the simplest version and layer from there. That my paragraphs land on a single declarative sentence rather than tail off. None of that lives in any style guide I could write down. It’s only legible from my corpus of text.

But my style changes dramatically from person to person — like how some people’s work voice jumps a whole octave when they speak.

Corpus-grounding only works if the corpus is consistent. And if you’re like me — your writing styles are different: one register on email, another on LinkedIn, a team voice, an external voice. Feed all of it to an AI and you get the average of your own mixed signals, which is also LinkedIn. Because I do ironically talk about leveraging and unlocking sometimes.

Here’s the thing I’ve learned: the work isn’t only on the AI side. It’s also on mine: figuring out what my voice actually is when I’m being myself. Or the version of me when I’m writing to that audience. Jung called it a persona. Claude called it a voice.

So I did some digging — and I let AI dig for me. I asked my AI: you have a voice style for me, but what if I have different voices? Can you do some research on how to manage these voices and what each one entails?

It came back with three useful frames. Mailchimp’s voice and tone guide as the canonical public example — one organization holding multiple registers across help docs, marketing, error messages, and support, all under a single identity. Nielsen Norman Group’s four tone-of-voice dimensions — formal/casual, serious/funny, respectful/irreverent, matter-of-fact/enthusiastic — useful axes for naming what you’re calibrating when you shift voice. And Erika Hall’s reframe in Conversational Design: voice is identity; tone is performance. Voice stays. Tone shifts to the room you’re in.

That helped me name what I was missing. I didn’t have a corpus problem — I had a tone-mapping problem. The voice was consistent; the tones I was deploying weren’t labeled or differentiated.

Then — I used my findings. The first is hand the AI a recent corpus of three to five paragraphs I wrote in the register I want it to match. Not my best paragraphs — my most representative ones. I want it calibrating on my typical, not my peak.

The second was name the anti-pattern explicitly. I don’t tell the AI what I sound like. I tell it what I don’t sound like. I don’t write thought leader or unlock potential or here’s the thing. I don’t open paragraphs with in my experience. I don’t end paragraphs by telling the reader what I just said. Specifics work; abstractions don’t. I let it do a first pass and then made sure it was right.

The third is edit the AI’s first draft as if I were editing my own. This is where the real shift happened. The AI’s draft was seventy percent there at best. The thirty percent that needs my hand is where my voice lived. If I accept the seventy percent, I’ve outsourced my voice. If I take the AI’s draft and rewrite the half-sentences that don’t sound like me, I’ve used the AI as scaffolding without giving up authorship.

Then I fed it back and told it to improve the style.

When I sit with practitioners wrestling with what AI means for their writing, I notice two reactions, often in the same person. One: oh my gosh, can you believe AI is doing this for me — doom and gloom for what writing used to feel like, beginning of the end of voice. Two: oh my gosh, I can’t believe AI can do this for me — genuine curiosity about possibility, about time recovered for the work that actually matters.

Both are right, and both are wrong. The doom is right that average AI prose is bland competence with a lack of critical thinking; wrong that there’s no escape. The excitement is right that there’s value available; wrong that it comes free.

Voice doesn’t get cheaper with AI. And neither does writing. What changes is where the work happens, and what comes back into the hours that used to be spent on the seventy percent.

Use the 70% as a starting point to react and refine. The question is whether we hand our voice over to the average — or do the work to keep it.