Why is AI-generated language so hard to understand?

Open original

An arXiv preprint finds that frontier language models produce text full of features that slow human processing, like unusual syntax and rare word combinations. The authors argue this stems from models being optimized for information density in math, code, and reasoning rather than for being easy to follow.

  • Compared with human writing, all nine tested frontier and near-frontier models use more inanimate subjects paired with animate verbs and more low-frequency n-grams.
  • Most models also rely more on long dependencies, stacked nouns, and reduced object relatives, especially in conversational text.
  • A proof-of-concept fine-tune under a length penalty pushed a model further away from human-like linguistic behavior, supporting the density explanation.

AI-generated text may be harder for people to read than human writing, and this paper tries to explain why. The authors start from psycholinguistic features known to slow human comprehension: object relative clauses, long syntactic dependencies, stacked nouns, and low-frequency word sequences (n-grams). They built a new corpus of AI-generated text across several domains and compared it against human writing and against paraphrases produced by round-trip translation, which serves as a control for meaning-preserving rewording. The headline result is that AI models use these hard-to-process features more often than the paraphrases do. Relative to human writing, all nine frontier and near-frontier models studied use more inanimate subjects paired with animate verbs and more low-frequency n-grams. Most also use longer dependencies and more stacked nouns, and most produce more reduced object relatives in conversational text. The authors also report substantial differences between model families, so this is not a uniform property of all LLMs. Their explanation is that human language evolved under pressure to communicate successfully with other people, while a growing share of model training targets math, code, and reasoning. Those domains reward information density, packing more content into fewer tokens, and that pressure can push prose toward compressed, harder-to-parse structures. To test this, the authors fine-tuned a language model with a length penalty and found that its output drifted away from human-like linguistic patterns. That is a proof of concept rather than proof of the mechanism in production models, and the paper is careful to frame it that way. Still, it gives a concrete, testable story for why a general sense of AI prose getting harder to follow might be real. The authors name the gap between human processing demands and LLM output as the core misalignment and suggest that remedies are possible, though the abstract does not spell out specific fixes. For anyone deciding whether to read the full paper, the useful parts are the feature-by-feature comparisons across models and the length-penalty experiment, which is the piece that ties the observed pattern to a plausible cause.

Ask about this article