Frontier language models systematically overuse six constructions that make their writing harder for people to process, according to a study from New York University and the University of Texas at Austin. The researchers compared output from 14 models with human conversations, technical discussions and student essays.
The sample included 120 conversations, 45 coding sessions and 980 essays. Models produced unusual word combinations, separated related sentence parts across long clauses and dropped connectors that would make relationships clearer. They also stacked nouns, turned nouns into uncommon verbs and assigned animate actions to abstract objects.
Smaller Llama and Olmo models were closer to human writing on the measured features. Within the Olmo family, deviation increased from the base model through instruction tuning to a reasoning version. The authors argue that training dominated by code, mathematics and verifiable rewards encourages dense, technically correct output without directly rewarding ease of reading.
That explanation is plausible but not proof that reasoning training alone causes every style difference. The study spans several tasks and models, yet readability also depends on prompts, audience and subject matter. A short-output penalty applied to a smaller model demonstrated that style can be shifted, but it overgeneralized the preference for brevity. The practical takeaway is to evaluate comprehension directly: correctness scores do not show whether a human reader can follow an answer easily.