A recent joint study by researchers at New York University and the University of Texas at Austin has identified six specific linguistic patterns that make text generated by frontier language models difficult for humans to read. According to reporting from The Decoder, these advanced systems systematically overuse complex structural habits that tax human parsing abilities. The investigation traced the root of this issue to how frontier models are trained, noting that training on code and math heavily rewards objective correctness while failing to prioritize human readability. Interestingly, the study observed that smaller language models tend to write in a style that is much closer to natural human communication.
The Cost of Code-Heavy Training
Large language models undergo extensive training phases that rely heavily on computer code and mathematical datasets to build logical reasoning capabilities. While this approach improves performance on technical benchmarks and factual problem-solving, it introduces unintended side effects into natural language generation. The rigid syntactic habits reinforced by code optimization do not translate well to prose meant for general human audiences. The published findings did not detail the exact methodology used to isolate the six patterns, nor did they publicly name the specific frontier models tested during the research.
Unanswered Questions and Next Steps
Several details regarding the study remain unknown. The announcement did not specify whether the identified patterns can be successfully mitigated through fine-tuning, nor did it state if major AI developers plan to adjust their training mixtures to address the issue. Observers will need to watch whether future model iterations incorporate readability metrics alongside traditional accuracy benchmarks.
As large language models become more deeply integrated into daily communication tools, understanding what makes machine-generated text fatiguing will be crucial for improving user experience.
Lead image: https://kaboompics.com/ / Pexels
