Why AI writing is so tiring to read

Someone on X asked a simple question this week. Why is AI writing so frustrating to read?

The usual answers are about telltale words, or about catching the machine in the act. That is not what was tiring me. I already know a lot of what I read now passed through a model. The problem is the sentence itself. I get to the end and still do not have the point.

Handwriting and a fountain pen

Handwriting and a fountain pen. Photo: Petar Milošević / Wikimedia Commons / CC BY-SA 4.0. Commons file.

The post was from Zhizhuo Zhou at Zonder Labs. The claim is blunt. AI writing often hides the main idea and invents awkward, overcomplicated phrasing. The sentences get harder than the thought requires.

They measured it. In their benchmark, AI articles used vague wording about 25% of the time. Human articles did it about 4% of the time. In Claude Code replies, vague wording jumped to about 60%, against about 7% in human posts on Stack Overflow. Hidden characters and actions, their phrase for when the real person and the real verb are not the subject and the verb of the sentence, showed up in about 9% of AI articles and about 1% of human ones. Empty statements were about 27% in AI articles and about 6% in human articles.

Those are their numbers, from a chart they published with the thread. I am not turning this into a model contest. Claude Code was the worst in their test. The everyday problem is broader. You can feel it in a work message, a summary, or a paragraph that looks polished and still asks you to do the unpacking.

The most tiring habit is a hidden subject. Their example turns “the duck walked up and asked” into “a request was made upon the duck’s arrival.” You have to dig the doer out of a noun. Then the words themselves go slightly wrong. Nouns get stacked into phrases nobody would say out loud. Their examples from real replies included things like “a label’s awareness entry” and “length floor drops a block.” I have read plenty of the same shape: a sentence I can parse, word by word, that still does not tell me what happened.

None of this is about whether the facts are right. It is not about a model sounding sure when it is guessing. It is not about spotting AI so you can stop reading. The prose can be confident, grammatical, even useful, and still wear you out because the idea is sitting one layer down.

If I have to read a sentence twice to find who did what, I ask for it again in plain English. I do the same to my own drafts. Put the person and the action up front. Say the thing once. If the writing still hides the point, it does not matter who wrote it.

More from Blog