Why AI text feels dense: research links readability to information density goals

A new study published on arXiv in October 2026 investigates why AI-generated text often feels difficult to parse. Researchers Thomas Hikaru Clark and two co-authors analyzed linguistic patterns across nine frontier and near-frontier language models to identify specific structural causes for this perceived complexity.

What happened

The researchers compiled a novel corpus of AI-generated text spanning diverse domains to test whether machine output differs systematically from human writing. They focused on psycholinguistic features known to increase cognitive load for human readers, such as object relative clauses, long syntactic dependencies, and low-frequency n-grams. The study compared these AI outputs against paraphrases generated through round-trip translation, which served as a baseline for simplified structure.

The analysis revealed that all nine tested models used more inanimate subjects paired with animate verbs and more low-frequency n-grams than human writers. Most models also exhibited longer syntactic dependencies, more stacked nouns, and more reduced object relatives in conversational contexts. The study noted substantial variation across different model families, indicating that architectural choices and training objectives influence these linguistic tendencies significantly.

To explain these findings, the authors proposed that while human language evolves for successful social communication, modern AI models are increasingly optimized for tasks like mathematics, coding, and logical reasoning. These domains prioritize information density over communicative ease. As a proof of concept, the team fine-tuned a language model with a length penalty, which forced the model to compress information. This adjustment caused the model’s output to diverge further from human-like linguistic behavior, supporting the hypothesis that pressure for density drives syntactic complexity.

How it works

The study relies on psycholinguistic metrics that measure how hard a sentence is for a human brain to process. For example, an object relative clause ("The book that the student read") is harder to parse than a subject relative clause ("The student who read the book"). Similarly, low-frequency n-grams refer to unusual combinations of words that do not commonly appear together in natural speech, requiring more mental effort to interpret.

The researchers hypothesize that the training objectives of large language models create a misalignment with human processing needs. When a model is trained to solve math problems or write efficient code, it learns to pack maximum meaning into minimum space. This "information density" goal conflicts with the redundancy and simpler structures humans use to ensure mutual understanding. By applying a length penalty during fine-tuning, the researchers artificially increased this pressure, demonstrating that the model responded by adopting even more complex and less human-like syntactic structures.

Key details

• The study analyzed nine frontier and near-frontier AI models across diverse domains.
• All tested models used more inanimate subjects with animate verbs compared to human writing.
• AI outputs contained significantly more low-frequency n-grams and longer syntactic dependencies.
• Round-trip translation paraphrases were used as a baseline for simpler linguistic structures.
• Fine-tuning a model with a length penalty increased divergence from human-like language.
• The authors attribute the complexity to optimization for information density in math and code.

Why it matters

For software developers and technical founders, this research explains why AI-generated documentation, code comments, or customer-facing content often feels stiff or overly dense. It is not merely a stylistic quirk but a structural outcome of how these models are optimized. Understanding that AI prioritizes information density helps teams anticipate where human editors must intervene to restore clarity and flow.

This misalignment has practical implications for product design. If an AI assistant is intended for casual conversation or educational support, its default tendency toward complex syntax may hinder user engagement. Teams building AI-native products need to account for this gap by implementing post-processing steps or fine-tuning strategies that prioritize readability over raw information compression, especially when the target audience is non-technical.

What you can do

• Audit AI-generated content for low-frequency n-grams and complex sentence structures before publication.
• Use round-trip translation tools as a quick heuristic to simplify overly dense AI output.
• Adjust system prompts to explicitly request simpler syntax and shorter sentences for user-facing text.
• Consider fine-tuning models with readability rewards if deploying custom LLMs for customer support.
• Train editorial teams to recognize specific markers like stacked nouns and inanimate subjects.
• Test AI outputs with real users to measure comprehension time, not just factual accuracy.