New York University and the University of Texas at Austin identified six linguistic patterns that advanced language models systematically use more frequently and that are harder for people to parse, according to The Decoder.
Authors link the problem to training on code and mathematical materials: there, correctness is rewarded but not necessarily clarity of exposition. According to the study description, smaller models, by contrast, write more like humans.
Practically, the takeaway is limited for now: the available material does not reveal the patterns themselves, the research methodology, or differences between specific models. Therefore, the publication signals a direction for testing the quality of responses rather than providing a ready universal readability criterion.
editorial commentary
Why it matters
A likely consequence is increased attention to readability as a separate quality metric for AI responses, not only their correctness. The next observable signal will be the publication of the full study with a list of patterns and the methodology for testing. Substantial uncertainty remains due to the absence of the primary text and data.