CASPER in the Machine: Insights into Character Variety in LLM-Generated Stories

This paper introduces the Character’s Portrayal Classifier (CASPER), a framework designed to analyze fictional characters through eight narratological dimensions: stylization vs. naturalism, coherence vs. incoherence, wholeness vs. fragmentariness, literalness vs. symbolism, complexity vs. simplicity, transparency vs. opacity, dynamism vs. staticism, and closure vs. openness. The researchers curated a dataset of 200 high-quality human-written short stories and 4,400 LLM-generated stories across four genres (Domestic, Romance, Science-Fiction/Fantasy, and Suspense/Thriller). By using LLMs as automated judges, the authors classified the protagonists of these stories to identify systematic differences in how AI and humans construct character identity.

The analysis reveals that LLMs do not mirror the character distribution found in human-written literature. Human authors frequently employ 'naturalistic' and 'fragmented' character portrayals, leaving room for reader interpretation and ambiguity. In contrast, LLMs heavily favor 'stylized' and 'coherent' characters, often over-explaining motivations and traits to ensure the narrative remains logical and straightforward. This suggests that while LLMs are highly proficient at generating consistent, readable characters, they struggle to replicate the nuanced, subtle, or intentionally disjointed characterizations often found in human creative writing.

As LLMs become standard tools for creative writing, it is essential to understand how they shape the narrative landscape. This research highlights a 'homogenization' effect where AI-generated stories lean toward specific, predictable character archetypes. For writers and developers, these findings provide a roadmap for identifying the limitations of current generative models, suggesting that human intervention is still required to introduce the depth, ambiguity, and artistic distortion that define compelling, non-formulaic literature.