July 9, 2026

Five patterns that make a text read like a machine

Readers rarely say "this is AI" because of a particular word. They say it because something in the shape of the text feels off, and they cannot name what. We can name it. Five things do most of the work.

1. Every sentence is the same length

This one is the loudest, and almost nobody notices they are hearing it. A language model picks the most probable next word, over and over, and the result settles into a comfortable middle: fifteen to twenty words, again, again, again. Read four such sentences aloud and you can feel the metronome.

Human writing does not do that. A person writes a forty-word sentence because the thought needed room, then a four-word one because it did not. Look at the paragraph above. Then look at this.

See?

The fix is not to shorten everything. It is to make the lengths disagree with each other.

2. Connectors announce every turn

"Furthermore." "Moreover." "Additionally." "It is important to note that." "In conclusion." These are the hinges of an essay written by someone who has been told to signpost, and models were trained on a great deal of that writing.

Good prose usually moves without announcing that it is moving. The relationship between two sentences comes from what they say, not from a word bolted on the front. Delete the connector and read again. If the passage still holds, the connector was scaffolding.

3. Symmetry, everywhere

"Not only X, but also Y." "It is not just a tool, it is a philosophy." "Whether you are a beginner or an expert." Three items in a list, always three, each about the same weight.

Balanced constructions are a real rhetorical device, and used once they land. Used in every third sentence they turn prose into wallpaper. Real thinking is lopsided: two examples, one of them longer than it deserved, and a third that turns out not to fit.

4. Every section closes itself

The tidy summarising sentence at the end of each block, the one that restates what you just read in slightly different words. Models produce these compulsively, because the training signal rewards completeness.

A reader who just read the paragraph does not need the paragraph again. Cut the last sentence of any section and check whether anything was lost. Usually nothing was.

5. Nothing is ever specific enough to be wrong

"Significantly improved performance." "A wide range of use cases." "Many users report." A model hedges by default, because a specific claim can be false and a vague one cannot.

Specificity is the cheapest way to sound human, and the one people skip. Four seconds, not "fast". Two of the eleven, not "several". A number nobody would bother inventing reads as lived experience, because that is usually what it is.

What this means in practice

None of these are about vocabulary, which is why swapping synonyms never works. They are about rhythm, structure and commitment. Youmanize rewrites for all five: it varies sentence length on purpose, strips announced transitions, breaks the balanced constructions, cuts the closing restatement, and leaves your specifics exactly as you wrote them. Numbers, names, dates and links are never touched.

Try it on something you wrote with a chatbot last week, and look at what gets cut.

Back to the blog