Study reveals LLMs fail to mimic human writing evolution, showing 'temporal flattening' across months and years of generated text

Researchers compared 412 human authors' writing across 2012-2024 with output from 3 major LLMs, analyzing 6,086 documents in academic, blog, and news domains

LLMs exhibit 'temporal flattening'—they cannot reproduce the natural evolution of writing style and cognitive patterns that humans develop over extended periods

While LLMs generate greater lexical diversity, they fail to show meaningful semantic and emotional drift that characterizes human longitudinal writing trajectories

Standard LLM generation treats each response as independent without memory, fundamentally limiting their ability to develop consistent evolving voice like human authors