Narrative Simplification in Human and AI Retelling

Human beings have long been regarded as "homo narrans," employing storytelling to comprehend and communicate reality. This study investigates whether contemporary large language models (LLMs) can approximate the depth and richness of human narrative abilities. Through an experiment comparing retellings by humans and various state-of-the-art LLMs on a dataset of man-made short stories, we examine how word count, vocabulary complexity, and readability evolve during the retelling process. We introduce the Flesch Reading Ease score as a quantitative measure of narrative simplicity. Our findings reveal an inverse correlation between narrative condensation and increased vocabulary density in both human and LLM retellings. While both humans and LLMs tend to condense stories, their approaches differ: humans generally maintain or improve readability, whereas LLMs, without specific instructions on persona-based retellings, often reduce readability. However, with appropriate prompts, LLMs can adjust their storytelling to more closely mimic human tendencies. We also identify a unique "adaptive retelling" ability in humans, who adjust readability based on the original story’s readability. These insights advance our understanding of the "AI narrans" concept, highlighting both the current limitations and the potential of LLMs in achieving human-like storytelling capabilities.