Blind test
Can you tell which one a human wrote?
Two texts on the same topic. One is real writing by a person; the other was generated from that person's measured style. Pick the human. Five rounds, no signup, and we publish the honest hit rate below.
5 answers so far — not enough data yet to publish an accuracy rate. We show the number once 30 answers are in.
Round 1 of 5 · Topic: Your first job
0/0 right
Your pick is recorded before the answer is revealed — that's how the published number stays honest.
Test your own writing
Paste something you wrote, then paste a version written in your voice by StyleMimic. We publish them side by side on a shareable page and count how often readers pick the real one.
Anything you publish here becomes readable by anyone with the link, so only use text you're happy to share.
How style detection actually works
When people say text "sounds like AI", they are almost never reacting to facts or grammar. They are reacting to habits. Default model output has an unusually even sentence length, few fragments, tidy connective phrases at the start of paragraphs, balanced clause structure, and a reluctance to be specific. Human writing is lumpier: a nine-word sentence next to a forty-word one, a dash where a comma would do, a detail nobody would invent.
Those habits are measurable, which is the whole basis of this test. Before generating anything, StyleMimic computes a fingerprint from your own samples: mean and median sentence length, how much that length varies, paragraph length, comma, dash, question and exclamation frequency per thousand characters, average word length, and vocabulary variety. It also picks two or three real extracts from your writing as examples, because a model imitates a sample far better than it follows a description of one.
Generation is then constrained by those numbers, and the result is measured again with the same code. That gives a style match score — how close the output's habits are to yours — and an originality score, which looks for five-word sequences shared between your samples and the output. High style match with low originality means the text merely recycled your own sentences, which is why we cap the headline score when that happens instead of celebrating a fake 100.
The pairs on this page are the same pipeline run on real writing from volunteers, with their permission. We keep the number public and unedited: there are not enough answers yet to publish a rate. Improving that number by rewriting prompts to game the test would defeat the point, so the pairs stay fixed and the answers stay as given.
Questions
How does the blind test work?
Each round shows two short texts on the same topic. One was written by a real person. The other was generated from that person's measured writing style — sentence rhythm, punctuation habits, word length. You pick the one you think a human wrote.
Is the accuracy number real?
Yes. It is computed from every answer given on this page, including the answers that make us look bad. Your pick is recorded before the correct answer is revealed, and replaying the page does not add new answers from the same session.
What does a low score mean?
If readers pick correctly around half the time, the generated text is statistically indistinguishable from the human original to a casual reader. Scores well above 50% mean the imitation is still detectable — which is exactly the number we are trying to improve.
Why can't I tell the difference?
Most AI detection cues are habits, not facts: uniform sentence length, no sentence fragments, tidy transitions, no personal specifics. When a model is constrained by a real person's measured habits and their own examples, those cues mostly disappear.
Can I do this with my own writing?
Yes. Paste 40 words or more of anything you wrote, and StyleMimic measures your style and drafts in it. You get a style match and an originality score on every result, so you can check the same thing we check here.