A new study investigates why AI-generated text has become increasingly difficult to read, examining nine frontier and near-frontier language models against human writing and machine-generated paraphrases.
Researchers analyzed psycholinguistic features known to make text harder for humans to process, including object relative clauses, long syntactic dependencies, and low-frequency word combinations. According to the paper, all nine frontier and near-frontier models tested use more inanimate subjects paired with animate verbs and more low-frequency n-grams compared to human writing. Most models also employ longer syntactic dependencies, more stacked nouns, and more reduced object relatives in conversational text.
When compared with paraphrases generated through round-trip translation—a technique that translates text and translates it back to check for comprehension—AI models still used these difficult linguistic features more frequently.
The researchers propose that this divergence stems from a fundamental difference in optimization goals. Human language evolved to prioritize successful communication and mutual understanding. AI language models, by contrast, are increasingly optimized for tasks like mathematical reasoning, code generation, and complex reasoning chains, where information density matters more than readability.
To test this hypothesis, the team fine-tuned a language model with a length penalty—essentially incentivizing it to convey information more concisely—and found that this constraint caused the model’s linguistic patterns to diverge further from human-like behavior, supporting their theory.
The study also documents substantial variation across different model families, suggesting that design choices and training objectives significantly influence how readable generated text becomes.
The findings highlight a misalignment between how humans process language and how large language models operate. The researchers suggest these insights could point toward possible fixes, potentially through training methods that better balance optimization for reasoning tasks with human readability.
Key facts
- All nine frontier and near-frontier models tested use more inanimate subjects with animate verbs and more low-frequency n-grams than human writing
- Most models employ longer syntactic dependencies, more stacked nouns, and more reduced object relatives in conversation
- AI text uses harder linguistic features more often than paraphrases generated through round-trip translation
- Models optimize for information density and reasoning rather than successful human communication