AI-isms: How to Spot AI-Generated Prose (and Clean Up Your Own)
You spot AI writing in fiction by its tics of regularity: lists of three, contrasts that oppose nothing ("it wasn't X, it was Y"), stacked adjectives, paragraphs that end on a moral, and a total absence of failed sentences. None of these markers is proof on its own; together, on one page, they are the signature of the statistical average that generated prose comes from. Here are twelve AI-isms, each with an example and its human version. They help you spot generated prose, and above all clean up your own, because most of them also live in writers who have never opened a chatbot.
Why generated prose has tics
A language model produces the most probable continuation. In a novel, the most probable continuation is the one thousands of texts have already written: the expected comparison, the ternary rhythm, the ending that explains. The result is clean and accident-free. But a voice is precisely a series of accidents you stand by. That is why a chatbot cannot write a novel in your place, as shown in can ChatGPT write a novel?.
The twelve markers
| Marker | Generated example (invented) | Human version |
|---|---|---|
| "The house was empty, silent, still." | "The house was empty." |
| "It wasn't fear, it was something else." | "She was afraid, and it annoyed her." |
| "An old, rusted, dilapidated, forgotten shed." | "A sheet-metal shed, rusted at the corners." |
| "And she understood that nothing would ever be the same." | (cut it; the scene is enough) |
| "A wave of sadness washed over her." | "She sat down on the crate." |
| "The rain fell, as if to underline the silence." | "It started to rain." |
| "Not only was he lying, but he knew it." | "He was lying. He knew it." |
| "The weight of the past pressed on her shoulders." | "She thought about March again." |
| "The dirty window bore witness to years of neglect." | "The window hadn't been washed in years." |
| "She opened her mouth, then closed it again." | "'Hi,' she said." |
| "He slipped out of the room." | "He left." |
| Every sentence runs 12 to 18 words. | A 30-word sentence, then a 3-word one. |
What these tics have in common
They fill. Each one adds words where the scene didn't ask for any: a third adjective, a comparison that explains what the reader had understood, a sentence that says what the previous one showed. Generated text has no intention; it compensates with texture. Beginner text often does the same thing, for a different reason: the fear that the reader won't understand.
That is why this list also serves writers who write everything themselves. The triplet, the closing moral and the emotion that washes over are in almost every first manuscript. That they have become AI markers gives you one more reason to hunt them down.
How to clean up your own
Three passes, on a finished chapter.
The paragraph-endings pass. Read only the last sentence of each paragraph. If it explains, concludes or generalizes ("and that is how", "she understood that"), cut it. Nine times out of ten, the paragraph ends better without it.
The rule-of-three pass. Look for lists of three adjectives, three verbs, three complements. Keep the most precise one, throw out the other two.
The abstractions pass. Look for abstract subjects (the weight, the silence, the sadness) performing an action. Replace them with what the character does. This pass joins the work on filter words and repetition described in how to spot repetition in your writing, and more broadly the method in how to improve your writing style in fiction.
In the Sagoræ studio, the AI Proofreader does this reading on a chapter marked as finished: it underlines a triplet, a paragraph ending on a moral, an "as if to", and explains why it is a tic. It does not rewrite the sentence; it lets you choose which of the three to keep. The built-in non-AI checker, free and unlimited, has already run the spelling pass before it.
C The Proofreader · Spelling, grammar, typography Pre-recorded demo
The Proofreader reviews a paragraph from chapter 3
The Proofreader
Five suggestions, to accept or reject one by one:
- grandmothers → grandmother's: possessive apostrophe.
- was creaking → were creaking: the subject is "the shutters".
- answer , → answer,: no space before the comma.
- did not answer, she was looking → did not answer; she was looking: two independent clauses, a comma splice.
- turning round : fine in British English; switch to "turning around" only if the rest of the novel is in American English.
I am leaving "at the slightest wind" alone: it is correct, and it is your rhythm.
The Proofreader
Understood: comma splices in dialogue-adjacent narration are added to your style exceptions for this novel, I will not flag them again. Four fixes applied, the paragraph is still yours.
Pre-recorded simulation. The built-in proofreader is free and unlimited; it suggests, the author accepts or rejects each fix.
Watch out for false positives
A triplet can be deliberate. A final sentence can be a punchline rather than a moral. No marker on this list proves a text was generated; they prove it has not yet been reread by someone looking for these tics. If the question behind your reading is "was this written with AI, and does it matter", it is treated in is using AI to write a novel cheating?.
Frequently asked questions
How do you recognize AI-written text? By the accumulation of regularity tics: triplets, hollow contrasts, stacked adjectives, a moral at the end of paragraphs, uniform sentence length, no failed sentences. One proves nothing; ten on a page do.
Are AI detectors reliable? No. They produce false positives on very clean human text and false negatives on generated text that was edited afterwards. A human reader who knows the markers remains more reliable.
Can a grammar checker spot AI-isms? A classic checker catches repetitions and some adverbs. A checker that reads style, like Sagoræ's AI Proofreader, also flags triplets and closing morals, without rewriting.
If you want a checker that shows you your tics without touching your sentences, Sagoræ offers a free trial, no card required.
✒️ Read next: Free Novel Writing Software: An Honest Comparison (Reedsy Studio, Novlr, Scribbook, WriteControl, bibisco, yWriter, Sagoræ) · Sudowrite Alternative for Writers Who Want to Write Their Own Words · WriteControl Review 2026: Features, Pricing and Who It's For
Write your novel in a real workshop: a focused editor, characters, story structure and AI coaches that guide you without writing for you. Start for free, no commitment.