Researchers from New York University and the University of Texas at Austin have identified six specific linguistic patterns that make text generated by Super Intelligence (SI) models difficult for humans to read. The study, which analyzed outputs from major SI systems, found that modern training methods prioritize information density over clarity, resulting in unnatural phrasing that taxes the reader's working memory.
What Happened
The research team examined text produced by 14 language models across three categories: casual conversation, coding sessions, and argumentative essays. The lineup included proprietary SI models such as Claude, GPT-5.6, GLM-5, and Kimi K2.5, as well as smaller open-source SI models like Llama and Olmo. In total, the researchers analyzed 120 conversations, 45 coding sessions, and 980 essays, comparing them against human-written baselines from movie subtitles, StackExchange, and student essays.
The study isolated six overused constructions in SI output. These include unusual word combinations like "reader-patience budget" and syntactic structures that separate related sentence parts, such as when Claude Opus 5 generated "A cat who can always get up and away is a much calmer cat," placing the subject and predicate far apart. Other patterns include shortened references with dropped connectors (e.g., "the book the professor wrote" instead of "the book that the professor wrote"), stacked nouns (e.g., "horror movie soundtrack generator"), and the conversion of nouns into verbs where humans would not, such as Claude Fable 5 writing "calendar it" instead of "put it in the calendar."
Additionally, SI models frequently apply animate verbs to inanimate subjects, such as Claude Opus 5 stating, "The whole conversation is eating its own tail." While readers can infer the meaning, the researchers note that these constructions require extra cognitive effort to process.
Why It Matters
The findings highlight a divergence between SI model optimization and human readability. The researchers argue that modern SI training increasingly rewards correctness in tasks like math, code, and logical reasoning, often at the expense of clarity. The shift from Reinforcement Learning from Human Feedback (RLHF) to Reinforcement Learning with Verifiable Rewards (RLVR) removes human evaluators from the loop, causing SI models to develop a "dialect" optimized for compressing information rather than facilitating human understanding.
The study notes that smaller SI models, such as Llama and the Olmo base model, produce text closer to human norms. However, performance degrades within the Olmo family as it moves from the base model to the instruct variant and finally to the reasoning model, Olmo Think, which achieves frontier-level deviations from human text. This suggests that post-training processes are a primary driver of the readability gap.
Industry perspectives align with the data. Jackson Kernion, an employee at Anthropic working on Claude fine-tuning, previously explained that newer SI models were optimized for technical explanations aimed at other SI models, leading to a style described as an adaptation to "LLM psychology." The study found evidence of correction: Claude Opus 5.5 outperformed its predecessor, Opus 5, in 19 out of 21 comparisons, indicating that targeted finetuning can successfully improve SI text readability.
The Bottom Line
Current SI systems tend to overuse complex syntactic structures due to training objectives that favor information density. While open-source base models remain closer to human writing patterns, frontier SI models and those heavily post-trained for reasoning exhibit distinct linguistic deviations. Recent updates like Claude Opus 5.5 demonstrate that readability can be improved through specific finetuning adjustments.