NYU/UT Austin Study: LLMs Overuse Six Working-Memory-Heavy Structures That Make Text Exhausting
xiaohu · x · 2026-10-27
Linguists at NYU and UT Austin compared 14 mainstream LLMs against human text to explain why AI writing feels exhausting even when it's correct.
Core finding: while human language balances speaker and listener effort, frontier models systematically overuse six structures that tax working memory:
- Unattested n-grams: odd coinages like "reader-patience budget"
- Long syntactic dependencies: subjects split far from their verbs by verbose modifiers
- Omitted relativizers: dropping "that" blurs clause boundaries and raises misreading
- Stacked nouns: e.g. "horror movie soundtrack generator", forcing readers to reverse-engineer modification
- Denomininal verbs: e.g. "calendar it" — compact for machines, jarring for humans
- Forced anthropomorphization of inanimate things: e.g. "the conversation is eating its own tail"
The paper includes a proof of concept: length-penalized fine-tuning shifts the model toward human-like linguistic behavior, mapping an axis of misalignment between LLM output and human processing demands.
More from AGI Musings
- Researcher: AI Tools Still Reward Domain Experts, Laypeople Can't Tell What Matters — IgorCarron · 2026-10-27
- Wolfram at ARC Prize: discovery means finding reducibility pockets in irreducible computation — mhmazur · 2026-10-27
- Yoshua Bengio discusses LawZero's misalignment work and UN-backed global AI governance — Yoshua_Bengio · 2026-10-27
- MIT's Andrew McAfee: we underestimated the scale, pace, and scope of tech progress — amcafee · 2026-10-27
- Model CoT laments mathematicians reduced to 'raw data' caretakers, fueling AI-debate discourse — aiamblichus · 2026-10-27
- Model trained on 40M faces predicts cancer survival from a selfie, stirring debate — TinfoilTricorn · 2026-10-27