The nine tells that make AI text obvious
Our published taxonomy sorts the habits that give a draft away into nine categories. Six are removals: most of what marks a draft can simply be deleted.
In short
The nine tells of AI text, in GenPolish’s published taxonomy, are stock openers, hedges, filler phrases, vague claims, clichés, padding, repeated shapes, the listicle reflex and over-signposting. Six of the nine are removals, so deleting them changes no meaning. The other three, vague claims, clichés and repeated shapes, change meaning when fixed, so each of those edits needs a person to check it.
A draft written by a model rarely fails on facts. It fails on register. The sentences are grammatical, evenly weighted, and hedged in the same places, and after two paragraphs a reader who edits for a living can name the source without being told. That pattern is what this taxonomy sets out to name.
Six of the nine categories are deletions: stock openers, hedges, filler phrases, padding, over-signposting and the listicle reflex. Nothing has to be rewritten to fix them. The sentence that remains is the sentence the writer meant, one clause shorter.
In today's fast-paced world, it is important to note that a stock opener is the cheapest tell to remove.
Three categories do change meaning: vague claims, clichés and repeated shapes. These are the ones we never apply silently. A vague claim is replaced with something specific, and something specific can be wrong, so the edit arrives with the phrase it replaced sitting next to it.
A tool that rewrites your argument without showing you is not an editor. It is a second author you did not hire.
What are the nine tells of AI text?
Can the tells be measured across many texts?
Style habits in machine text can be measured, but only in bulk. Kobak and colleagues analysed more than 15 million PubMed abstracts from 2010 to 2024 and found that the excess vocabulary of 2024 was made almost entirely of style words: of 379 excess style words, 66 per cent were verbs and 14 per cent adjectives. Delves appeared 28 times more often than its earlier trend predicted, underscores nearly 14 times.
From that excess the authors estimate that at least 13.5 per cent of 2024 abstracts were processed with a model. They also name the limit of the method, which works on the corpus and “cannot identify individual abstracts that may have been processed by an LLM.” The taxonomy takes the same position. It marks phrasing a reader can see and fix, and draws no conclusion about who wrote the page.
Is the human score an AI detector?
The GenPolish human score is a reading of register against the nine categories. It is not a detector, it does not identify a model, and it cannot tell you whether a passage will be flagged somewhere else. Text a person wrote in stock business phrasing scores badly, and that is the correct result: the phrasing is the problem either way.
The practical version of all this is short. Cut the opener. Cut the hedge. Replace one vague benefit with a number you can defend. Most drafts improve more from those three moves than from another round of prompting.
Questions
Which words give AI text away?
Style words give AI text away across a corpus, not in one document. An analysis of more than 15 million PubMed abstracts found that delves appeared 28 times more often in 2024 than its earlier trend predicted, with underscores and showcasing close behind. The authors estimate at least 13.5 per cent of 2024 abstracts were processed with a model, and state that the method cannot identify individual abstracts.
Does a low human score mean a machine wrote the text?
A low GenPolish human score does not mean a machine wrote the text. The score counts phrasings from the nine published categories per hundred words and makes no claim about authorship. A person writing in stock business phrasing scores badly, which is the intended result: the phrasing is the problem whoever produced it, and each mark is a phrase you can inspect and reject.
What is the fastest way to make AI text less obvious?
The fastest fix for obvious AI text is removal, not rewriting. Cut the stock opener, cut the hedge, and replace one vague benefit with a number you can defend. Six of the nine categories in the GenPolish taxonomy are deletions, so most drafts improve with nothing added, and the sentence that remains is the one the writer meant, one clause shorter.
Which AI writing fixes can change the meaning of a sentence?
Three of the nine tells change meaning when fixed: vague claims, clichés and repeated shapes. A vague claim is replaced with something specific, and something specific can be wrong. A cliché gives way to a plain verb that may say less or more. GenPolish never applies these three silently, and each edit arrives next to the phrase it replaced, for the writer to accept or refuse.
Sources
- How Claude’s text watermarking worksAnthropic ·
- The nine categories, with definitionsGenPolish ·
- Delving into LLM-assisted writing in biomedical publications through excess vocabularyScience Advances ·