StoryScope: Investigating Idiosyncrasies in AI Fiction (arxiv.org)

A new study introduces StoryScope, a pipeline designed to investigate the narrative structures of AI-generated fiction. As AI-written content proliferates, the project emphasizes distinguishing these works from human-authored stories based on discourse-level features rather than superficial stylistic signals. StoryScope analyzes a substantial parallel corpus of over 61,000 stories produced by five large language models (LLMs) and human authors, extracting 304 narrative features across ten dimensions. The research reveals that AI narratives tend to over-explain themes and maintain simpler, more linear plots, contrasting with the moral complexity and temporal richness found in human narratives. Notably, the model achieves 93.2% accuracy in detecting human versus AI authorship using narrative features alone, retaining over 97% of the accuracy achieved with stylistic signals. This study not only enhances the understanding of authorship in the age of AI but also suggests that fundamental differences in narrative construction can effectively identify the origins of fictional works, highlighting the unique storytelling characteristics emblematic of human creativity versus machine generation.