← All posts

Technology · July 22, 2026 · 2 min read

Voice Matching Is Not Style Transfer

Tone sliders and 'write like a pirate' prompts miss what a personal voice actually is. What it takes for AI to sound like a specific person.

Ask a chatbot to "write this in a casual, confident tone" and you'll get something casual and confident — and completely generic. That's style transfer: mapping text onto a broad register. A personal voice is a different object entirely, and the difference explains why most AI-written posts are recognizable at a glance.

Voice is a bundle of small, stable habits

What makes your writing yours is not "tone." It's dozens of measurable, mostly unconscious regularities:

A tone slider controls maybe one of these. The rest default to the model's own averages, and the averages are exactly what readers detect as "AI-sounding."

Why examples beat instructions

Here's the thing about those habits: you can't describe them, because you don't know you have them. No one can fill out a brand questionnaire that captures their sentence-length distribution. But your published posts encode all of it implicitly. That's why modern voice systems learn from examples — real posts you actually shipped — rather than from adjectives you picked off a list. The writing is the spec.

The hard part is what to keep

Matching a voice isn't imitating everything. Some patterns in your old posts are voice; some are noise; some are habits you'd rather drop. A good system has to separate the signal ("always opens mid-story") from the incidental ("mentioned baseball twice in March"). This is also why voice systems improve with feedback: every edit you make to a draft is a label telling the system which patterns were actually you.

A test you can run today

Paste three of your posts into any AI tool and ask for a fourth on a new topic. Then check the details: Did it match your paragraph lengths or default to uniform blocks? Did it use punctuation you never use? Did it end with a question you'd never ask? The gap between "sounds nice" and "sounds like me" is where this technology is actually competing.