Essay
Why AI Can't Write Like You
AI writes clean, competent text all day, but getting it to sound like one specific person is a much harder problem.
AI can write pretty much anything you ask it to, but it has a much harder time sounding like one specific person. I wanted to understand why, so in October I built Voicework, a local app that learns how you write from your own writing and the corrections you choose to keep. Most of what I learned came from watching it get things wrong.
First you have to know what you actually wrote
Obviously, teaching it my voice meant starting with my own writing. Figuring out what actually counted as mine was harder than it sounds, though. I went through the 61 documents it found on my computer and labeled each one: about 40 were mine, 8 were a mix of me and AI, 9 were basically AI, and 3 weren’t mine at all. For the hard cases, my rule was simple: if it was written before ChatGPT came out, it’s 100% mine.
That cleanup mattered way more than I’d expected. I’d written a long technical paper with a lot of AI help, but it sounded nothing like me. A few days later, I realized the prompts I type to AI tools were another problem. They’re basically instructions, not really my voice, and they’d been teaching it the wrong thing the whole time.
It kept making things generic
The first big failure came when I asked it to tighten up my writing. It treated phrases I actually use, like “pay for pixels,” as too inefficient to keep and cut them, flattening my writing into generic technical prose.
I tried it on excerpts from one of my articles and preferred my original in 2 of the first 3 comparisons. A rewrite would make me think “this feels very AI,” or I’d see a sentence slip straight into typical AI rhetoric. In one case, the only change I even wanted was a double dash – to show it’s not the AI dash. I retuned it to keep my wording and explain things the way I do, and after that I picked its version in 22 of 22 excerpts.
It missed the small stuff
Testing it on messages I’d held back from training showed what it was missing from my texts. None of its versions caught the things that make them sound like me: “bro,” “ngl,” “much love bro,” short paragraphs broken over a few lines, and an all-caps “LET’S DO IT 🚀.”
Later, I did a blind check and actually picked the version with no voice guidance twice. It came down to one tiny detail each time: a sign-off in one, and an all-caps ending with an emoji in the other. With the sign-off, I’d never say “love you,” only “much love.” Measuring sentence length and word choice alone can’t catch either detail.
Other mistakes mattered a lot more than whether it sounded like me. In one version, “once I’m settled” became “once you’re settled.” Another rewrite changed “This case is deliberately limited” to “I deliberately limited,” which gave me credit for a decision I never made. That’s honestly what bothers me more than anything about style: it’s now saying something that isn’t true.
The context labels were brittle
I don’t write to a friend the way I write in a bio or a work message, so it’s mostly a good thing that Voicework keeps those contexts separate. But changing the wording showed me how brittle the labels were: with the exact context from my article edits, it found 22 relevant lessons. When I called the same thing a “technical article,” it found zero.
I have plenty of texts and work messages, but nothing I’d written for a public website. That explained why it matched only one lesson and zero samples when I tried it on this website’s copy. It flagged all six of its own changes, and I didn’t end up applying any of them – honestly the right call, since it didn’t have anything to go on.
Using another AI as the judge
I also tried having a model judge the rewrites, and it always ranked the version with no voice guidance last. It could at least tell that something had changed, but choosing between the voiced versions was another matter: it disagreed with me 3 out of 3 times in one check, then agreed only 3 out of 6 times in the next. In the end, I was still the only one who could really judge whether something sounded like me.
What actually helped
The most useful input came from corrections I chose to remember, each tied to the situation it came from, and from blind comparisons where I didn’t know which version was which. Samples shouldn’t be the only way to teach it.
Being honest about how little we’d actually proven helped too. After one round of checks, the calibrated version had won 2 comparisons and lost 1, with another where I didn’t like either version. That’s not enough to claim an accuracy percentage, so the app shows which checks have actually been done instead of a confidence score.
I ran this piece through Voicework too, using the lessons from my own edits. Funny enough, the first draft still came out too polished for me, so I went back and loosened it up.