What PlainSpeak cannot do
PlainSpeak is a finishing tool. This page lists what it does not do, so you know when to rely on it and when to use your own judgement.
The scan
- It reads English only. Its phrase lists, patterns and reading ease formula are all English. A draft in another language finds few or no tells and can score 100 however generated it reads, so the score, the re-scan and the guard that keeps your original when a rewrite reads worse all mean nothing there. The word count and the phrase lists read English letters only, so accented and non-Latin words are split or skipped: a word with an accented letter is read without that letter, sometimes as two words, and a word in another script is not counted at all. The rewrite itself keeps the language you give it.
- It reads wording and a few patterns, not meaning. It matches stock phrases and a handful of shapes, such as sentences all the same length or each clause picking up the last one's word. A fluent machine draft that avoids them can score clean.
- The phrase list is an editing heuristic, not a validated reader result. We chose its words and phrases as an editor would, and nobody has tested whether readers notice each one, or how much.
- It is not an AI detector. A clean score says the wording has few of the phrases and patterns on our list. It makes no claim about what a detection service would say, and it cannot prove who or what wrote a piece.
- Style is described, not scored. Passive voice, hedging, long sentences and the rest of How it reads are shown but never change the score, because a contract, a paper and a social post are all right to differ.
- The score is an editing policy, not a probability. Its weights, ceilings, 250-word floor and bands are choices, not measured constants, and it has never been calibrated on labelled writing. A score of 100 can still have tells to fix.
- Style measures are approximate. Passive voice and abstract nouns are found by word patterns, not grammar, and reading ease is a formula over sentence and word length. They describe surface style; they cannot tell whether a reader understands the piece or whether it is accurate.
- Short texts give weak numbers. Under about 120 words the change-from-source number after a rewrite is marked low-confidence.
Your writing baseline
- It describes; it does not predict. The rank says where a draft's opening sits among your own pieces at one length. It is not a calibrated probability, and it does not promise that a set share of your future writing will be flagged. With five pieces, one new piece of your own can easily land past all of them.
- It compares the opening only. The rest of a long piece is scanned but not compared with your writing.
- It is not a model of your writing. It chooses which general edits apply and stops the rewrite from forcing your draft back to an AI source's register. Your voice and lens guide the rewrite; PlainSpeak does not imitate you.
- Its settings were developed on one author's writing. The comparison lengths, the five-piece minimum and the flag thresholds were worked out on a small private set of one writer's pre-AI articles, kept for testing and never shipped in the app. They have not been checked across many writers, forms or languages.
The rewrite
- Models vary. Some shorten text when asked not to, some add em dashes, some keep the habits they were asked to drop. PlainSpeak catches the worst cases and keeps your original, but the fix is often another model.
- It needs your own AI key, always. In the desktop app every feature is free for 7 days, counted from the moment you select Start the trial, and nothing that uses AI runs before that press. After the 7 days, rewriting, fact checks and source checks need a licence; scanning, the highlights and editing by hand stay free, before, during and after the trial. See Your free trial.
- It cannot show whether a hidden watermark survived, and it does not claim to. We tested that question offline, with the open SynthID-Text scheme (
eval/watermark-check.md): across three runs of 18 marked drafts, a Repurpose's output lowered the detector's signal in every piece of all three, and left the mark above a measured null in 1 of 18, 12 of 18 and 18 of 18 of them. That is our own key and an open scheme, not Google's detector, so it bounds the claim from below rather than proving anything about a vendor. No surface of the app claims a watermark was removed. - The revert check counts wording tells only. When a rewrite has clearly more stock phrases, jargon, clichés, connectors or antitheses than your draft, PlainSpeak keeps your text. It does not count em dashes, rhythm or how the piece closes, so a rewrite can pass the check and still read flatter. See The rewrite read more generated than your text.
- Four-gram numbers are not word counts. One changed word changes up to four four-word sequences, so "output four-grams changed" is not the share of words edited.
The fact check
- It is mechanical. It compares numbers, dates, names, links and quotations between your original and the rewrite. It cannot tell you whether a claim still means what it meant.
- Find sources needs an OpenRouter key, and only marks a source as checked when the page really contains the claim's figures.
The audit
- It needs OpenRouter, and runs on Rewrite and Repurpose only.
- It is two more opinions, not proof. It is two models' judgment of the wording: models share blind spots, and their agreement does not show who wrote a piece or that its facts are right. After a few rounds what is left is often a matter of taste.
- One reply is not an audit. If one model fails or replies off-format, the verdict is INCOMPLETE, not PASS.
The app
- Windows and Mac only. There is no phone or tablet version.
- Link import works only for pages anyone can open. Text files can be up to 1 MB. Word files can be up to 10 MB, and the text in them up to 1 MB.
- Preview, Copy and Download .docx format the Markdown PlainSpeak understands: headings, bold, italic, lists, quotes, links, code, tables and dividing lines such as
---. Images show as you typed them. A link keeps its address only when it starts withhttp:,https:ormailto:; any other link keeps its text. - Side by side needs a window about 1,160px wide or wider. In a narrower one the original and the rewrite are stacked. Below about 1,530px the AI scan panel shows as a strip at the right that opens as a drawer.
- Very long pieces. Above about 60,000 characters (roughly 10,000 words) the editors and Preview stop showing scan highlights, and linked scroll follows by position rather than by paragraph. The scan itself still runs.
- Word files come in as text: images, footnotes, comments, text boxes and dividing lines are left out, and tables become plain paragraphs. The old
.docformat does not open; save it as.docxin Word first.
The evidence behind the numbers on this page, including what we measured with our own watermark key, is in Protocol.