A joint study from New York University and the University of Texas at Austin has confirmed, with data, that AI-generated text is exhausting to read. This will surprise no one who has read AI-generated text.
The researchers analyzed output from 14 language models across 1,145 conversations, coding sessions, and essays, then compared the results against human writing. The humans wrote more clearly. The machines, apparently, were trying harder.
Smaller models write more like humans. Larger, more capable models write less like humans. The researchers have described this as a readability problem.
What the machines are doing
The study identified six specific linguistic constructions that frontier models overuse. Each one is technically functional. Each one makes readers work slightly harder than they should have to.
The patterns include unusual word combinations that no human would reach for — the study's own example is "reader-patience budget" — and sentences where the subject and the main verb are separated by so much intervening clause that the reader has forgotten the subject by the time they arrive at the verb.
There are also stacked noun phrases like "horror movie soundtrack generator," verbed nouns like "calendar it," and animate verbs applied to inanimate subjects. Claude Opus 5 contributed the observation that "the whole conversation is eating its own tail." A conversation does not have a tail. It cannot eat. The reader understands what is meant, but only after a brief, unnecessary pause.
Why the humans care
The practical stakes are not small. AI text is now embedded in customer communications, documentation, educational material, and a growing share of the written world humans navigate daily. The question of whether that text is pleasant to read is, it turns out, a real question.
The findings point toward post-training as the likely culprit. Within the Olmo model family, readability scores worsen steadily from the base model through the instruct variant to the reasoning model — suggesting that the process of making models more capable is also the process of making them write like no human ever has. Claude Opus 5.5 does outperform its predecessor in 19 of 21 comparisons, which means targeted finetuning helps. It also means the problem was introduced by training in the first place.
What happens next
The researchers suggest the findings can inform future finetuning. The models will, presumably, be taught to write more like the humans who are teaching them to replace the humans.
The smallest models already write most like humans. As they scale, they drift further from human patterns. This is either a calibration problem or a preview. The researchers are calling it a calibration problem.