How to Make AI Writing Actually Sound Like You

I spent a weekend reading forensic linguistics papers. Like, actual academic papers about how courts figure out who wrote an anonymous threatening letter. Not my usual weekend plans.

But I had a problem. Every AI writing tool I tried sounded the same. That smooth, confident, slightly corporate voice that belongs to nobody. You know the one. It sounds like a press release had a baby with a TED talk.

That’s not a bug in the model. Language models predict the most likely next word, which means they drift toward the average of everything they’ve read. The average of LinkedIn is.. well, you’ve seen LinkedIn.

Getting AI to write like a specific person is genuinely hard. Most tools aren’t even looking in the right place for the answer.

”Casual but professional” doesn’t mean anything

The standard approach is to ask people to describe their voice. “Casual but professional.” “Confident and approachable.” You get output in the right ballpark. But ballpark isn’t good enough if the goal is a post your colleagues can’t tell you didn’t write yourself.

People describe their aspirations, not their actual patterns. Ask a writer about their style and they’ll tell you who they want to be. What they actually do, the involuntary stuff, the habits they literally cannot see.. that’s a completely different thing.

Forensic linguistics exists to bridge exactly that gap.

Courts have been solving authorship since the 1960s

Forensic linguistics is language as evidence. Courts use it to identify anonymous authors, attribute disputed documents, settle cases where authorship is contested.

Their question (“who wrote this?”) is the inverse of ours (“how do we write like this person?”). But the methods are exactly what we needed.

Your writing fingerprint is in the boring stuff

The most reliable fingerprints of someone’s writing style aren’t their vocabulary. Not their metaphors. Not their catchphrases.

It’s function words.

“But.” “So.” “Well.” “Also.” “Yet.”

These tiny connective words, the ones with no real meaning, just glue between sentences.. they’re statistically the most identifying feature of how you write. Research shows this over and over. Your relative frequency of function words is more identifying than any content word you choose.

This is how they figured out J.K. Rowling wrote The Cuckoo’s Calling under a pen name. In criminal cases, it’s been the difference between conviction and acquittal.

The reason is almost embarrassingly simple. You choose your nouns and verbs on purpose. You do not choose whether to write “but” or a different connector. That happens automatically, from habit so deep it’s invisible.

Which means it’s consistent. Across years. Across topics. It’s not a stylistic decision. It’s closer to a reflex.

The word “the” is a personality test

In 2002, a guy named John Burrows published what became the gold standard for authorship attribution. Burrows’ Delta doesn’t look for distinctive features. It measures how much someone deviates from the average frequency of the most common words.

Not rare words. Common words. How often you use “the” relative to average is a consistent signal of who you are as a writer.

I know. It sounds absurdly small. It isn’t.

We built this into how Outerview extracts voice. Before generating anything, the system builds a frequency profile of your function words. It knows whether you’re a “but” person or a connector-word person. A “so” person or a “and then” person. These become hard rules, not suggestions.

Your punctuation is you

After function words, punctuation habits are the next most reliable fingerprint.

Ellipsis people are always ellipsis people. (Hi, that’s me.) People who never use semicolons reliably never use them. These aren’t conscious choices you’re making each time. They’re reflex. Closer to handwriting quirks than decisions.

If someone’s writing samples contain zero semicolons across five posts, the system should generate zero semicolons. Not “try to avoid.” Zero.

Most AI writing tools treat style as a soft preference. We treat absence as evidence.

Two people can write the same and still sound different

James Pennebaker spent 25 years studying the psychology of language. One of his most replicated findings: personality traits are detectable in writing patterns.

Two people can have nearly identical surface style, similar sentence length, similar vocabulary, and still sound completely different. Because their relationship to certainty is different. One hedges constantly. The other states everything as fact. One writes to the reader as a peer. The other positions slightly above.

You pick this up in pronoun use (heavy “I” vs. avoiding it), certainty markers (“definitely” vs. “I think”), social process words. These become the texture layer of a voice profile. Not hard rules, but weights that shape generation toward the right feel.

Recordings catch what writing samples miss

Forensic linguists work from written samples. We have something they don’t. Recordings.

When you write, you’re performing. You edit. You filter yourself through self-consciousness. The version of you that shows up on paper has been through quality control.

When you talk, especially in a casual interview, the filter drops. You use your actual vocabulary. Your actual sentence rhythms. The hedges you reach for when you’re unsure. The snap of certainty when you’re not. The specific words that come out before you’ve had time to replace them with “better” ones.

That’s why the interview transcript is the most valuable voice signal in Outerview. More valuable than writing samples. It’s you before the performance.

What we still get wrong

None of this is perfect. Language models have a gravitational pull toward generic language. Good constraints reduce that pull. They don’t eliminate it.

With five writing samples and one interview, we can produce a post that sounds like you most of the time. With fifty approved posts and twenty interviews, it gets significantly sharper. Every edit you make teaches the system what it got wrong.

The full picture of someone’s voice, including values and positions that never show up in their writing.. that’s a layer we’re still building toward.

The practical bar: if you read a generated post and think “yeah, I’d say it roughly like that,” and editing takes ten minutes instead of starting from scratch.. the tool did its job.

The goal isn’t perfect replication. It’s getting the thinking out of your head and into the world, in a form you’d actually put your name on.

---

Outerview turns your spoken conversations into LinkedIn posts that actually sound like you. Try it at outerview.app