SettledResearch
Paper finds frontier models write with harder-to-read language than humans do
An arXiv paper (Clark, Martin, Mahowald, submitted Oct. 21) finds nine frontier and near-frontier models use more difficult-to-process features than human writing, such as low-frequency n-grams and inanimate subjects with animate verbs. Difficulty varies by model family. The authors suggest training for math and code drives it; a length-penalty fine-tune moved output away from human-like patterns.
- HEAT
- OUTLETS
- 0
- HACKER NEWS
- 197 pts
- FIRST SEEN
- LAST SIGNAL
Primary source
arXivarxiv.org/abs/2610.29162Coverage · 1 article, oldest first
Saturday, October 24, 2026
arXivPrimary197 pts on HN
Why is AI generated language so hard to understand?