New arXiv paper: all 9 frontier models write 'harder' language, info density is why
_onionesque · x · 2026-10-23
A UT Austin preprint by Thomas Hikaru Clark, Matias Martin and Kyle Mahowald asks why AI-generated text is getting harder to understand. Using a novel corpus of AI text, they find all nine frontier and near-frontier models use more inanimate subjects with animate verbs and more low-frequency n-grams than humans, with most also using longer dependencies, stacked nouns, and reduced object relatives; model families vary substantially. Their hypothesis: human language is optimized for communication, while LLM output is increasingly optimized for math, code, and reasoning, favoring information density. Fine-tuning under a length penalty diverges from human-like linguistic behavior, supporting the idea and pointing to remedies.
Related event: Preprint Explains Why AI-Generated Language Feels So Weird(3 posts)→
More from Research
- Michael Levin's new slide: mind encloses matter, not the other way around — cephaloform · 2026-10-23
- OpenTac from JHU and MIT auto-designs 3D tactile skins for robot hands — anand_bhattad · 2026-10-23
- Stanford's Anshul Kundaje calls BioReason's variant effect prediction evals seriously flawed — anshulkundaje · 2026-10-23
- Science cover: 3-agent AI nearly doubles physician diagnostic accuracy at lower cost — EricTopol · 2026-10-23
- 3-agent AI nearly doubles diagnostic accuracy vs physicians at lower cost in 600+ NEJM/ER cases — EricTopol · 2026-10-23
- DeepMind paper: model bidding auction hits 99.2% of optimal routing without knowing which model is best — rohanpaul_ai · 2026-10-23