How the score works
This page explains the AI scan panel: the score, the tells it counts, the style it describes, and the voice number after a rewrite.
The score
The score starts at 100 and drops for each kind of tell PlainSpeak finds. It counts wording a reader may notice, nothing else. The list of words and patterns is an editing heuristic, not a validated reader result.
| Score | Band |
|---|---|
| 95 to 100 | Low tell density |
| 80 to 94 | A few tells |
| 55 to 79 | Tell-heavy |
| Under 55 | Dense with tells |
The score is a weighted density: wording tells are counted per thousand words, so a long piece is not punished for being long. Very short pieces are scored as if they were 250 words, so one stray word in a short note does not sink it. Each kind of tell has a ceiling, so one habit cannot sink the whole score.
The weights, the ceilings, the 250-word floor, the band cut-offs and the rates that put stock wording over tolerance are editing choices, not measured constants. The score is not the chance that a piece was written by AI, and it says nothing about who wrote it.
100 does not mean nothing to look at. In a long piece, two stock phrases can round the score back to 100. The chips still show their exact count, and they are counted below the line.
100 does not mean a person wrote it. The score reads wording, and current models avoid most stock phrases, so their drafts often score 100. It cannot tell who wrote a piece.
What is scored
These are the tells. Each shows as a chip with its count; select a chip to see only those in your text. Groups with no hits are folded away: select Show N groups at zero to see them too.
Wording a reader may notice
| Tell | What it is |
|---|---|
| AI-sounding | Phrases models overuse: "delve", "tapestry", "unlock the potential". A reading claim, not an authorship one: words careful human writers use as freely, like "leverage" and "innovation", count as corporate speak instead. |
| Corporate speak | Business jargon: "leverage", "synergy", "circle back". |
| Clichés | Worn-out metaphors: "game changer", "at the end of the day". |
| Connector spam | Too many signposts: "Furthermore", "Moreover", "In conclusion". |
| Antithesis | The "it's not X, it's Y" move, used again and again. |
How the draft is built
| Tell | What it is |
|---|---|
| Clause chaining | Each clause picks up the last one's word: "builds trust, and trust builds loyalty". |
| Uniform rhythm | Every sentence, or every paragraph, about the same length. |
| Tidy close | The last paragraph opens by summing up: "Ultimately", "In short". |
What readers believe is a tell
| Tell | What it is |
|---|---|
| Em-dash pile | More than three em dashes. Many good writers use them, so this is weak evidence, but readers judge it anyway. |
Not scored, but worth a look
| Tell | What it is |
|---|---|
| Rule of three | Lists of three: "X, Y and Z". One is fine. In a piece of 300 words or more, 14 or more per 1,000 words goes over tolerance and a rewrite is asked to vary them, without changing the score. With a writing baseline, your own range decides instead. |
| Vague attribution | "Studies show", "experts say": a claim that names nobody a reader could check. A count and highlights only, for you to name the source or say it as your own view. It never counts toward a fix or the score. |
What the scan skips
Some text in a piece is not your wording, so the scan leaves it out of every tell, every style measure and the word count the score is based on:
- Quotations in double quotes, straight or curly. A quotation can run over several lines, but not past a blank line.
- Code, both inline code in backticks and fenced code blocks. A fence with no closing fence runs to the end of the piece.
- Web addresses, and the address part of a Markdown link. The link's own words are yours, so they are scanned.
In a rewrite, a quotation or a piece of code is skipped only when your original has it too, so a rewrite cannot hide new tells by putting them in quotes. A quotation in the rewrite that is part of a longer quotation in your original also counts as yours. Headings and bold text are scanned like the rest of your copy. Single quotes do not mark a quotation, because the same mark is an apostrophe. The word count in the editor still counts every word, since a rewrite is sent all of them.
What is described, not scored
How it reads shows style measures that never change the score, because a contract, a paper and a social post are all right to differ: Hedging, Adverb pile, Passive voice, Rule of three, Abstract nouns, Long sentences, Repeated openers and Serial comma.
Some are flagged when there is a real pile: for example, long sentences making up over a quarter of a piece. Hedging and Adverb pile are never flagged from a count alone, because neither is a sign of who wrote a piece: they show a count, and What to watch next time says when there are enough to look at. Passive voice is plain writing advice too, and is not flagged from a count alone: current models use it about as often as people or less, so saying who did it makes a sentence clearer, not less like a model. With a writing baseline, these measures are judged against your own usual range instead.
These are approximate surface measures. Passive voice and abstract nouns are found by word patterns, not by parsing grammar, so "the station" counts as an abstract noun and some passives are missed. Reading ease is a formula over sentence and word length; it does not measure whether a reader understands the piece.
Over tolerance
Below the chips, Over tolerance lists every group with enough hits to be worth fixing. Stock wording needs a rate as well as a count, so two stock phrases in a long piece stay below the line, where the same two in a short piece do not. Below the line counts tells that appear, but not often enough to be listed. A rewrite is still asked to tidy a stock phrase below the line; Fix these N is not. On the Current copy tab, Fix these N fixes just the groups over tolerance, N being how many those are. See Rewrite your copy.
The tiles
| Tile | What it shows |
|---|---|
| Average sentence | Average words per sentence. |
| Long sentences | The share of sentences over 26 words. |
| Reading ease | The Flesch reading ease score, with its band in words. Higher is easier; 60 to 70 is plain English. |
| Reading time | Reading time in minutes, at 200 words a minute. |
Select the i on a tile to see what it measures.
Original and Current copy
Original scans the draft you brought in. Current copy scans the rewrite, and updates as you edit it. After a rewrite, the panel switches to Current copy by itself.
Change from source
After a Polish or Rewrite, PlainSpeak measures how far the rewrite's style moved from the draft you brought in, across 14 measures such as sentence length, rhythm, reading ease, contractions and hedging. It shows as change from source with a verdict and a percentage:
- held: the rewrite stayed close to the source's style.
- drifted: it moved noticeably.
- very different: it moved a long way.
This is movement from the source, not distance from how you write. When the source came from an AI model, staying close to it keeps the model's habits. With a writing baseline, a measure that moved but whose opening sits inside your usual range does not count. The label then reads source change outside your range and names the opening it measured, such as the first 300 words. The comparison with your own writing is the separate pair of lines under the score, such as "Further from your writing than 4 of your 8 pieces".
The held and drifted lines are editing choices, not validated limits: Polish allows 12%, Rewrite 36%. Under about 120 words the number is marked (short text), because a few words can move it a long way.
Repurpose sets out to change the style, so it is never marked held or drifted. With a Repurpose lens, it shows target difference: how far the output sits from the lens, as a measurement, not a pass or fail. Without a lens it shows no target.
Looks like
The Looks like button names the kind of piece PlainSpeak thinks you pasted: long-form, social or conversational. It decides from headings, length, hashtags and mentions, and whether the text talks to the reader. Select it to see why, or to correct it. The kind sets how the Mass audience voice shapes a rewrite, which template Repurpose suggests, and whether your writing baseline applies.