Study links AI-generated text’s processing difficulty to information density goals
Researchers find frontier models use more low-frequency words, longer dependencies, and inanimate subjects with animate verbs compared to human writing, suggesting optimization for density over clarity.
1 source · cross-referenced
- A study of nine frontier and near-frontier AI models finds their text uses more low-frequency words, longer syntactic dependencies, and inanimate subjects with animate verbs than human writing.
- Researchers propose AI-generated text is optimized for information density rather than successful communication, diverging from human language patterns.
- The paper introduces a corpus of AI-generated text across diverse domains and demonstrates model-family variation in linguistic features.
- A fine-tuning experiment with a length penalty further shows divergence from human-like linguistic behavior.
A new study published on arXiv examines why AI-generated language often feels harder for humans to understand than human-authored text. The authors analyze psycholinguistic features known to correlate with processing difficulty—such as object relative clauses, long syntactic dependencies, and low-frequency n-grams—across a novel corpus of AI-generated text spanning multiple domains. They report that AI models use these difficult features more frequently than paraphrases produced by round-trip translation. When compared directly to human writing, the study finds that all nine frontier and near-frontier models examined exhibit patterns associated with higher processing load: more inanimate subjects paired with animate verbs, greater use of low-frequency n-grams, and in most cases, longer dependencies, more stacked nouns, and reduced object relatives in conversational contexts.
The authors propose a hypothesis to explain this divergence: while human language is optimized for successful communication, AI-generated text is increasingly optimized for objectives like math, code, and reasoning, where the pressure favors information density over clarity. To test this, they fine-tuned a language model with a length penalty, which led the model’s output to diverge further from human-like linguistic behavior. The study also documents substantial variation in these features across different model families, suggesting that design choices and training objectives shape how models generate language.
The paper contributes a new corpus of AI-generated text and a set of quantitative comparisons against human baselines, highlighting axes of misalignment between human processing demands and LLM outputs. The authors suggest their findings point toward possible remedies, including alignment techniques that explicitly account for human readability.
- Nov 11, 2026 · The Verge — AI
Google DeepMind unveils AlphaGenome 2 to predict genetic variant effects on cell types
Trust79 - Nov 11, 2026 · Google DeepMind — Blog
Google DeepMind unveils AlphaGenome 2 to predict DNA variant effects in specific cell types
Trust84 - Nov 11, 2026 · arXiv cs.AI
Researchers demonstrate auditable human-agent teams for autonomous synchrotron X-ray experiments
Trust79