Study finds AI writing leans on sentence patterns that are harder to read

Researchers at NYU and UT Austin compared text from 14 models, including Claude, GPT-5.6, GLM 5 and Kimi K2.5, with human writing and with plain paraphrases of the same content. Compared with human writing, all nine frontier and near-frontier models used more lifeless things doing what people do (as in "this asymmetry bites") and more rare word combinations, patterns linked to harder reading. The authors suspect that training for maths, code and reasoning rewards packing in information rather than being understood; when they trained a small model with a penalty for length, it shifted to shorter, less human phrasing. Claude Opus 5.5 scored lower than Opus 5 on 19 of 21 difficulty comparisons, so newer models can improve. This is an early preprint, English only, with a small set of prompts. The Hacker News thread (144 comments) has no consensus. Four top-level commenters say they have no trouble reading AI text. The biggest sub-thread, 39 replies, starts from a technical writer who thinks AI answers are tuned to raters' first impressions of good writing; one reply there finds AI text hard because it is information sparse, not dense, which cuts against the paper. Four top-level comments and one reply share fixes, such as an "ELI5" follow-up, "summarize again without the jargon", or asking for an "8th grade reading level". Sorted take: plain-language instructions are cheap to try, but the thread shows readers differ, so test drafts on the people who will actually read them.