A study by researchers at New York University and the University of Texas at Austin reveals that leading AI models, including Claude Opus 5 and GPT-5.6, overgenerate stylistic patterns that impair readability. The researchers analyzed output from 14 language models across 120 conversations, 45 coding sessions, and 980 essays, comparing them against human-written baseline datasets.
The study identified six distinct patterns that tax working memory, including weird word combinations, distant predicates, dropped connectors, stacked nouns, unnecessary verbification, and animate verbs applied to inanimate objects. The authors attribute this emergent dialect to post-training techniques like Reinforcement Learning with Verifiable Rewards (RLVR), which optimize for logic and correctness rather than human readability.
While newer iterations like Claude Opus 5.5 and targeted penalization during training showed potential readability improvements, current reasoning-focused models consistently diverge from natural human sentence structures.
Why it matters
Post-training techniques like RLVR optimize for accuracy over human readability, creating an awkward synthetic dialect.
Founders building customer-facing text products must implement finetuning or prompt engineering to mitigate cognitive load on users.
Readability metrics show that base models often mirror human text better than heavily post-trained reasoning variants.
Source: the-decoder.com



