AI BriefingStudy Finds AI Models Use More Difficult Psycholinguistic Features Than Humans
·2026.10.25 03:04
Key point
Nine frontier and near-frontier models use difficult psycholinguistic features more often than human writing, attributed to optimization for non-communication goals.
Details
- AI-generated text is harder for humans to process because models use difficult psycholinguistic features (e.g., object relative clauses, long syntactic dependencies, low-frequency n-grams) more often than human writing or round-trip translation paraphrases.
- All nine frontier and near-frontier models use more inanimate subjects with animate verbs and more low-frequency n-grams, while most also use longer dependencies, more stacked nouns, and more reduced object relatives in conversation.
- The study attributes this shift to optimization for goals like math, code, and reasoning, which prioritize information density over successful communication.
- A proof-of-concept fine-tuning experiment using a length penalty demonstrates that this optimization pressure leads to a divergence from human-like linguistic behavior.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
NeurIPS 2026 paper proposes sampling coherent but cognitively unavailable research directions
Audible uses controllable TTS and BookLLM to improve audiobook character differentiation