Here’s a thing I see constantly.
Someone hears AI is the future. They open ChatGPT, Claude, Gemini, Bloomberry, whatever the app of the week is. They paste in three LinkedIn posts they wrote in 15 minutes on a Wednesday. They type: “Train on my voice. Write me a post.”
Out comes something that sounds like a slightly-more-coherent version of themselves on a bad day. They shrug. “Good enough.” They post it.
Next week, they do it again.
And again.
And again.
Six months later, their entire public voice has drifted toward the blandest possible version of themselves. They cannot figure out why engagement is down. Why their audience feels distant. Why nobody’s really tracking them anymore.
They blame the algorithm. They switch tools. They swap from ChatGPT to Claude. They try a fancier “voice cloning” app. Nothing gets better.
Because the problem was never the tool.
The problem is that the voice they trained the AI on was already a diluted version of themselves.
Brands Don’t Die in a Blaze. They Die in a Blur.
This is how brands die in the AI era. Not with a bang. With a thousand slightly-generic posts.
Every time you publish something that’s 85% you and 15% AI-ambient-beige, you shift your audience’s mental model of you a little. Do that a hundred times in a row and you’re a different person on the feed. You don’t sound like you anymore. You sound like “a coach.” Or “a consultant.” Or “a woman who left corporate and now talks about leadership.”
The worst part? You can’t feel it happening in real time. It’s drift. Slow, steady, frictionless drift.
Which is why the fix is not to find a better AI.
The fix is to have a real voice worth training AI on in the first place.
What “Voice” Actually Is (Not Adjectives)
Most brand voice docs are useless. They look like this:
- Professional
- Approachable
- Bold but warm
- Confident but humble
- Authoritative yet conversational
Great. Now describe the taste of water.
Voice is not adjectives. Voice is specifics.
Voice is the actual words you use and the ones you refuse to use. It’s your sentence rhythm. The jokes you would never tell a client but absolutely think. The metaphors you reach for. The grammatical rules you happily break. The hill you’d die on in a meeting. The thing you always say at dinner that makes your friends laugh. The way you start an email when you like someone versus the way you start one when you don’t.
A real voice doc is 2-4 pages. It has:
- 20 phrases you genuinely say
- 10 phrases you’d never say (these are just as important)
- 5 opinions you would defend in public
- 3 structural tics in how you write (do you use parentheticals? dashes? fragments?)
- Examples. Lots of them. Full paragraphs, not bullet points.
If the AI’s output sounds generic, it’s not because AI is dumb. It’s because the voice input was thin.
Why Most “AI Voice Training” Fails
You know what most people feed AI when told to “train on your voice”?
- A handful of LinkedIn posts
- Maybe a website About page
- Maybe a bio
Those are the most polished, most self-conscious, most filtered versions of how you write. You were performing when you wrote them. You hedged. You softened. You business-ified.
So when AI averages them, it produces the most performed, most hedged, most business-ified version of you possible. Dressed up in your cadence, drained of your actual personality.
Garbage in, nicely-formatted garbage out.
The better inputs? The stuff you never publish:
- Long DMs to friends where you actually explain your opinion
- Voice memos you made to yourself brainstorming out loud
- Angry drafts you never sent
- Emails to clients where you were genuinely helpful
- Transcripts of talks, podcasts, or interviews
- Slack messages where you forgot to be polite and were just right
That is where your voice actually lives.
That is what you should be feeding the AI.
The Way I Actually Build an AI Copywriter
When I build Radical Branding™ for a client, we spend serious time on voice before any AI gets involved.
Not a surface-level intake form. Real work:
- A long interview where I get them talking instead of writing
- Collection of their actual writing across contexts (polished and messy)
- A curated library of sentences and phrases that are unmistakably theirs
- A list of words, phrases, and patterns they hate
- Hot opinions with context
- Running jokes and recurring references
Then – and only then – I build Obie. Their custom AI copywriter. Trained on the real voice doc. Fed real samples. Tuned against the “never say this” list.
The difference is embarrassing.
Generic AI output sounds like every coach on LinkedIn. Obie’s output sounds like my client. Like, their spouse-would-recognize-it sounds like my client. Because we didn’t train on the filtered version – we trained on the actual human.
The AI didn’t invent voice. The strategy did. The AI is just the megaphone.
What You Can Do This Week (Without Hiring Anyone)
You don’t need a 75-day brand build to start fixing this.
Here is a small, doable homework assignment:
1. Pull 10 pieces of “real” writing. Not your website. Real. Long emails to a colleague. DMs with a friend where you explained something. A draft you never published. Paste them all into one doc.
2. Highlight the sentences that sound the most like you. Not the best ones. The most *you* ones. Maybe it’s a cussy line. Maybe it’s a weird metaphor. Maybe it’s an unusually short paragraph that lands like a slap. Mark them.
3. Write a “never say” list. 10 phrases you refuse to let out of your mouth. “Elevate your journey.” “Unlock your potential.” “In today’s fast-paced world.” Whatever makes you cringe.
4. Put all of that into your AI’s system prompt. Every single conversation. “Here are 20 real sentences of mine. Here are 10 phrases I never use. Write in this voice or don’t bother.”
5. Refuse to publish anything that fails the voice test. If your best friend read it and would not say “oh yeah, she wrote that” – rewrite. Or don’t post.
That is how you stop training AI on a diluted version of yourself.
That is how you use AI to sound more like you, not less.
Voice is the moat. Do not let a robot take it from you – especially a robot you fed yourself.
If you want to hear what a strong, specific voice actually sounds like in action, the guests on Curveball the Podcast are a parade of them. Former executives, artists, healers, and hell-raisers who sound like absolutely no one else. That is what we’re protecting.