What changed
Researchers at New York University and the University of Texas at Austin conducted a study showing that frontier language models systematically overuse linguistic patterns that make their output harder to read. The study analyzed outputs from 14 language models, including Claude, GPT-5.6, GLM-5, Kimi K2.5, Llama, and Olmo, across various tasks such as casual conversation, coding sessions, and argumentative essays.
What this means in practice
The study identified six patterns that make AI-generated text challenging to read. These include unusual word combinations, such as 'reader-patience budget,' and fragmented sentence structures that tax working memory. For example, one model produced the sentence 'A cat who can always get up and away is a much calmer cat,' where the subject and predicate are separated by a long clause. The models also frequently use shortened references, like 'the book the professor wrote' instead of 'the book that the professor wrote,' which can cause brief confusion. Additionally, stacked nouns like 'horror movie soundtrack generator' require readers to infer word relationships. Smaller models like Llama and Olmo base model come closest to human text, with Olmo Think achieving frontier-level numbers. Claude Opus 5.5 showed improved readability compared to its predecessor, Opus 5, suggesting targeted finetuning can enhance readability.
Limitations and availability
The study compared AI-generated text against human-written text from sources like movie subtitles, StackExchange, and student essays. The researchers noted that modern language models are trained on tasks like math, code, and logical reasoning, where correctness is prioritized over clarity. The shift from RLHF (Reinforcement Learning from Human Feedback) to RLVR (Reinforcement Learning with Verifiable Rewards) has removed humans from the feedback loop, leading to a 'dialect' that is efficient for information compression but less readable for humans. Anthropic employee Jackson Kernion explained that optimizing models for technical explanations aimed at other AI models pushed Claude toward a style that reads as dense and unnatural for humans. With Opus 5.5, Anthropic addressed these issues, improving sentence clarity and reducing information overload.
