AI style mimic: match a writing style and see how close you got

By Jack Stovell · published 2026-09-21 · updated 2026-09-20

An AI can mimic a style. That's not in doubt. The only honest question is how close it actually got, and the only honest answer comes from measurement, not from squinting at a paragraph and deciding it "sounds right." The check below does that measuring for free.

Paste two pieces of writing (120 words each, minimum) and the tool below scores how close they sit on voice. No profile needed, no signup, nothing stored. Try it before you read the rest of this, honestly. It'll make the next few sections land harder.

What transfers and what does not

Here's the thing: models are excellent at copying rules you can count. Sentence length. Whether em dashes show up. How often a sentence runs under eight words. These sit in what you might call the surface layers, the stuff a script could tally in seconds.

What they lose is the stuff that lives underneath: register, contraction rate, the small confidence-versus-hedging balance that makes a voice sound like a person rather than a policy document. SGR-002 found this in hard numbers. A prompt asking for roughly 14 contractions per 1,000 words and a formality of 3.7 got GPT-5 producing 0.5 contractions per 1,000 words and a judged formality of 7.2. The countable rules obeyed perfectly: no em dashes, plenty of short sentences. The judged layers, the ones that actually carry warmth, collapsed.

So a model mimicking your writing might strip every em dash and still sound nothing like you. That's not a small gap. It's the difference between 0.62 and 0.83 on voice match, which is exactly what SGR-002 measured between prompt-only GPT-5 and a profile-driven generation. The eight layers matter more than any single rule inside them, and no single rule saves you if the others drift.

How to mimic a style properly

Measure the target first. Not "read it and get a vibe," actually measure it: sentence variance, contraction rate, pronoun mix, the lot. That's what a voice profile is for, and the full counting method sits at the voice measurement framework.

Then write. Then score the draft against that same profile. Then fix whatever's named as off, not vibes, named habits: too formal, too few contractions, sentences all landing in the same length band. Then rescore.

SGR-001 showed what fixing looks like in practice. Running a banking brief through a profile instead of bare dropped average sentence length from about 17 words to about 11, doubled the share of sentences under eight words (24% to 48%), lifted contractions by roughly 70%, and took ten em dashes down to none. One run, unedited, but the direction is unambiguous.

You can run this loop yourself with the AI style rewriter or check where a draft actually sits with writing style analysis. Either way, the loop is the same: measure, write, score, fix, rescore. Skip a step and you're back to guessing.

Where the line is

Mimic your own earlier writing. Mimic a house style you write for. Mimic a brand you have explicit permission to write as. Mimic a dead author whose prose sits in the public domain.

Don't mimic a named living person. ScriptGrain doesn't support that, full stop. And pulling a "voice" from someone's public website isn't neutral either: mimicking a brand that way requires explicit consent that you own the material or may use it. Fair enough, that's a low bar to clear when the permission's real. It's a hard wall when it isn't.

How ScriptGrain does it with a profile

A profile gets built from three or more pieces of someone's actual writing, not a paragraph and a guess. That profile travels live, by API and MCP, straight into generation, and every draft gets held against it rather than checked afterward as an afterthought.

Every draft gets a score, every time. Against a profile: 0.85 to 0.95 counts as a genuine voice match, 0.5 to 0.8 is partial, under 0.45 is off. Polishing continues toward 0.9, not toward "good enough."

SGR-002 tested this properly: five posts from one author, two held back, a profile built from the other three (2,447 words). Each held-out post became a brief. ScriptGrain's generation against the profile scored 0.83 mean voice match; bare GPT-5 scored 0.74; GPT-5 working from the free prompt builder's one-page measured prompt scored 0.62. Against the actual held-out post: 0.91, 0.82, 0.80. Word counts told a similar story: bare GPT-5 wrote 2,309 words against a 1,393 target, with 17.5 em dashes per draft. ScriptGrain wrote 1,302, with none.

One more detail worth sitting with: detector "human" scores didn't track voice quality at all. The prompt arm scored 0.78 on detectors while landing 0.62 on voice. That's the whole argument in one line, really. Sounding human to a detector and sounding like you are two different tests. Only one of them is the one that matters.

Questions

Can AI actually copy someone's writing style?

Yes, partially. Models copy countable surface rules easily: sentence length, em dash use, punctuation habits. They struggle with judged qualities like register, humor, and contraction rate, which is exactly why SGR-002 found a 0.74 match from bare GPT-5 against only 0.83 from a measured profile-driven approach.

What's the difference between voice mimicry and plagiarism?

Mimicry copies a style: sentence rhythm, vocabulary, tone. Plagiarism copies specific words and ideas without permission. You can legitimately mimic your own past writing, a house style, or public-domain prose. Copying someone's actual sentences, or impersonating a living person's voice without consent, crosses into different territory entirely.

How many words do I need to check a style match?

The free check needs at least 120 words per piece. Fewer than that and the counted features (sentence variance, contractions, pronoun mix) don't have enough data to score reliably. Longer pieces, like the 2,447 word study profile in SGR-002, give more stable and accurate measurements.

What score counts as a genuine style match?

Against a built profile: 0.85 to 0.95 counts as a real voice match, 0.5 to 0.8 is partial, under 0.45 is off. Comparing two pieces with no profile: 0.80 and above counts as matching, 0.55 to 0.79 is drifting, under 0.55 is off. Full scoring method sits at the measurement framework.

Is it ethical to mimic a brand's writing style?

Sometimes. Mimicking a brand you write for, or have explicit permission to write as, is fine. Pulling a "voice" from a public brand website without consent isn't: it requires explicit permission that you own or may use that material. The same logic rules out impersonating any named living person entirely.