NYU/UT Austin study: all 9 frontier models write text harder to process than human writing
windx0303 · x · 2026-11-21
A paper by Thomas Hikaru Clark (NYU), Matias B. Martin, and Kyle Mahowald (UT Austin) asks why AI-generated language is hard to understand.
- Analyzes psycholinguistic features linked to human processing difficulty (object relatives, long syntactic dependencies, low-frequency n-grams) across a novel multi-domain corpus of AI text
- Compared to human writing, all 9 frontier and near-frontier models use more inanimate subjects with animate verbs and more low-frequency n-grams; most also show longer dependencies, stacked nouns, and reduced object relatives in conversation, with substantial variation across model families
- Hypothesis: human language is optimized for successful communication, while AI text is increasingly optimized for goals like math, code, and reasoning that reward information density; a length-penalty fine-tuning proof-of-concept reproduces the divergence
- Highlights axes of misalignment between human and LLM linguistic processing demands and points to remedies
More from AGI Musings
- Gary Marcus mocks 'AI psychosis' hype over AI-solved math problems with satirical test — GaryMarcus · 2026-11-21
- 250,000-page Riemann proof is more like an iPhone than magic: the AI math comprehension debate — nptacek · 2026-11-21
- After AI Cracked Navier–Stokes, Millennium Problems Fell Fast—Why Did Tao Miss It? — burny_tech · 2026-11-21
- Neuroscientists warn top journals are full of 'polished garbage' systems neuroscience papers — KordingLab · 2026-11-21
- Forecasters put 85% odds on AI solving another Millennium Problem by 2027 — NathanpmYoung · 2026-11-21
- Kurzweil's next book: 'The Singularity Came and Went' — PeterDiamandis · 2026-11-21