Paste a draft. Plainspoken shows you the exact sentences that make it read as machine-written and why, then rewrites it to match how you actually write. Nothing gets invented on the way.
How we know: we measured 75 properties of prose across 2,048 human documents and 1,560 machine ones and kept the ones that actually separate the two. Not a wordlist. Not a vibe. The measurement is public.
Orange marks are constructions and words that are rare in human writing and common in machine writing. They are what a reader notices. The score above is driven mostly by things no single sentence shows, like how evenly your sentence lengths are spread.
Two rounds, three candidates each. Every candidate is checked for invented numbers, invented names and dropped figures before it can win. Usually 10–20 seconds. Limit 10 a day per address while we are testing.
Every tool that promises to "humanize" text hands your draft to a general-purpose model with the instruction to sound human. That makes text more detectable, not less, and it routinely introduces typos and invents citations. We measured it.
Plainspoken does three things differently. It measures the draft first, against a specific person's real writing, and turns the gap into arithmetic: three of your eleven sentences must be under eight words; remove these four words; cut the three-item lists. It generates several candidates and scores each one again, keeping only what actually moved. And it refuses any candidate that adds a number, adds a name, or drops a figure that was in your source. If nothing passes, you get your original back and a reason.
The score runs in your browser with no model behind it: 66 measured properties and a set of coefficients that fit in 6 KB. That is why it is instant and why it can tell you which sentence, where a percentage from a black box cannot.
Outbound email, follow-ups, LinkedIn posts, proposals, anything where a reader deciding you used a model costs you the reply. Readers who believe a message is machine-written trust it less, and that penalty lands before any detector is involved. The point is not to fool software. It is to write like yourself, at volume, without saying anything false.
If you are trying to pass off machine text as your own work for a class or a publication that forbids it, this is the wrong tool and we would rather you did not. Plainspoken does not hide anything and emits an edit trail. It makes AI-assisted writing sound like you; it does not make it not AI-assisted.
Scoring and diagnostics, unlimited. Rewrites, 10 a day. No account. This is a test and it will stay free until the measurements against outside detectors are published.
Your own voice profiles fitted from your real writing, unlimited rewrites, an API, and the evidence trail. Not charging until it earns it. Leave an address if you want to know when.
Testimonials. Plainspoken is new and there are none yet, and we are not going to make any up. There is no "undetectable" claim either. The only efficacy number we have is our own instrument grading text that was optimised against our own instrument, which proves the loop works and nothing else. The independent test against named detectors is the next thing we build, and no claim goes on this page before it exists.