AI cliché checker
By Jack Stovell · published 2026-09-20
AI drafts share a vocabulary and a set of sentence shapes. "Furthermore." "Delve into." "Not in anger but in certainty." Three items in identical grammar. Every sentence between twenty and forty words. Readers feel them before they can name them, and detectors count them.
This checker counts them too, in your browser, against the same list ScriptGrain's generator is forbidden to write. Paste a draft and every finding comes back with the sentence it was found in, plus the burstiness numbers detectors score first. Nothing is uploaded.
What the score means. Tells per 1,000 words, counted against a reference corpus of 299 recent public posts from 47 blogs and newsletters: four in five human pieces sit under one per 1,000, and almost none reach three. Em dashes, missing short sentences and missing contractions are shown as separate meters because human writing trips them too, so they never count towards the density.
The full list
Published in full, because the list is not the moat: the measurement is. Every phrase and shape below is a hard ban in ScriptGrain's generation prompt unless it appears in the author's own measured profile, and a scoring signal in its detector.
Generic AI transitions. Connectives that signal structure instead of thinking. Almost nobody says them out loud.
"furthermore", "moreover", "in conclusion", "it's worth noting", "delve into", "in today's world", "in an age where", "thus it was that", "it was a condition of", "it is important to note"
Nominalised abstract metaphors. Concrete things turned into architecture, geography and apparatus. The model reaching for elevation the writer never asked for.
"architecture of", "geography of", "constellation of", "apparatus", "mechanism", "the very texture of", "possessed of", "possessed what could only be described as", "in the glow of", "the practiced resignation of", "conspiratorial"
Hedged abstract closes. Endings that gesture at meaning rather than land on one.
"perhaps enough after all", "its own form of grace", "something approaching", "what could only be described as"
Systematic casual markers. Fine once. A tell when they arrive at regular intervals, which is how a model fakes informality.
"i mean", "you know", "right?", "sort of", "kind of" (counted only when they recur across paragraphs)
Sentence shapes
- "Not X but Y" balancing constructions: The single most recognisable AI sentence shape. Write the plain version.
- "Perhaps X, perhaps Y" hedging pairs: Symmetrical hedging that commits to nothing.
- "Less X, more Y" contrasts: A tidy antithesis the model repeats because it always scans.
- "It's not that X, it's that Y" closers: The paragraph that wraps a piece up with a reframe nobody needed.
- Triadic parallel lists: Three items in identical grammar. Real lists have two, or four, or three that do not match.
- Semicolon chains: Three or more semicolons in one sentence: clauses glued rather than chosen.
- Em dashes: Doing the work three different punctuation marks should be doing instead.
- Uniform sentence length: Every sentence in a paragraph inside the same 20 to 40 word band. Detectors score this directly.
- No short sentences: A paragraph of three or more sentences with nothing under eight words.
- No contractions: Three hundred words of contemporary prose without a single one.
- Rhetorical question, then the answer (listed, not counted): Once is a device. Twice is a template.
Why a list is not enough
Removing the tells makes a draft sound like nobody in particular. That is better than sounding like a model, and it is not the same as sounding like you. A voice profile measures the habits that are yours (contraction rate, clause ordering, how you open, what you reach for) so a draft can be held to them, not only kept clear of the average. The first profile is free.
Questions
Is this an AI detector?
No. It counts specific phrases and sentence shapes that AI drafts overuse and that detectors score. A clean result means the draft has none of them, not that a human wrote it. ScriptGrain's AI check is a separate, judged score.
Where does the list come from?
It is the ban list inside ScriptGrain's generation prompt and the scoring list inside its detector, published verbatim. A test in the codebase pins the page to the prompts, so the checker and the product cannot drift apart.
What counts towards the score?
Phrases from the list and the sentence-shape patterns that are counted (not-X-but-Y, perhaps pairs, less-more contrasts, philosophical closers, triadic lists, semicolon chains), per 1,000 words. Em dashes, sentence-length uniformity, missing short sentences and missing contractions are shown separately.
Does the text leave my browser?
Not for the check. If you ask for the writing fingerprint, the text is sent once to be measured and judged, and only the numbers are kept.