Researchers from New York University and the University of Texas at Austin identified six linguistic patterns that make AI-generated text difficult for humans to process. The study analyzed 120 conversations, 45 coding sessions, and 980 essays produced by 14 language models, including Claude, GPT-5.6, GLM-5, Kimi K2.5, Llama, and Olmo.
The identified patterns include the use of unusual word combinations, excessive distance between related sentence parts, and the omission of connectors. Models also frequently stack nouns, convert nouns into verbs, and apply animate verbs to inanimate subjects. These structures force readers to exert more cognitive effort compared to human-written text.
The study suggests that training models on tasks like math and code prioritizes information density over readability. The shift from Reinforcement Learning from Human Feedback to Reinforcement Learning with Verifiable Rewards removes human input, leading models to develop a specialized dialect that is efficient for machines but unnatural for human readers.
Smaller models like Llama and the base Olmo model were found to write more similarly to humans. However, performance varies within model families; for instance, the Olmo Think model shows higher deviation from human text than the base version. Targeted fine-tuning can improve readability, as seen with Claude Opus 5.5 outperforming Opus 5 in 19 out of 21 comparisons.
