What GenPolish does, and what it refuses to do

Why GenPolish exists

Drafts written with an AI model share the same habits: stock openers, stacked hedges, lists where a paragraph would do, and a handful of words that turn up far more often than in human writing. Readers notice them, even when they cannot say why.

GenPolish removes those habits without taking the draft out of your hands. It proposes each change on its own, shows it next to your original, and applies only what you accept.

The list of what we change is public

The engine sorts every edit into nine change categories, and each one is defined in writing. Seven are found by matching a turn of phrase. Two — repeated shapes and the reflex to turn prose into a list — can only be seen in the shape of the text, so they are named without being rewritten.

Four things are preserved by default: numbers, URLs, citations and Markdown. The engine does not rewrite an argument.

There is no proprietary score you cannot inspect. A score you cannot argue with is a score you cannot check, and the taxonomy is published for exactly that reason.

What we remove, and what we refuse to promise

Rewriting text does not prove authorship, and it does not guarantee any detector result. We do not sell a way around detectors, and we do not claim one.

The register score describes how machine-like a text reads against our published taxonomy. That is all it describes. It is not a probability that a machine wrote something, and it should not be used as one.

What a scanner can see, and what it cannot

Our tools recognise eight invisible characters. They are real, and they do turn up in text that has passed through a model, an editor, a PDF or a web page. Finding one points to a text-handling step, not to an author. Every rewrite removes five of them, and leaves the three joiners alone because Arabic, Indic scripts and emoji need them.

A watermark is a different thing. It is a statistical bias applied while a model picks its words, so no scanner can read it from the text alone, and cleaning characters does not touch it. Replacing the words does, because the words are what it is made of, and replacing them is what a rewrite does. We report what came out and what share of your words changed. We do not claim what a detector would conclude, because the keys that read these marks are not public.

The watermark database records what each of five providers actually does, what the evidence is, and when it was last checked.

Five tools that never send your text

Five utilities inspect, clean, compare or count text entirely in your browser. They need no account. The text you put in them never leaves your device, and what you do in them is never measured, even if you accept analytics: only opening the page counts toward our daily arrival totals.

The AI rewriter is the exception, and it is explicit: it sends a draft to the server when you ask for a rewrite, and only then.

How we check what we say

Each provider record cites primary sources and carries the date it was last checked. Each article cites what it relies on, in links that are followed rather than devalued.

Automated tests keep that policy enforceable. A character named in the prose must exist in the engine. A claim that a provider marks its text must carry a primary source. A date must be a real one. And the only measurement of our own that the site reports is the study of AI tells published on the blog, with its method, its sample sizes and its limits.

What is not built yet

Two parts of the product are announced but not built: an HTTP API, and team accounts. Each says so on its own screen, and neither offers a control that pretends to work.