StoryScope Identifies Distinct AI Fiction Across LLMs
TL;DR. Researchers developed StoryScope, a pipeline distinguishing AI-generated stories from human narratives based on discourse-level features, not just style. - StoryScope achieved 93.2% accuracy in human vs. AI detection and 68.4% for six-way LLM attribution. - AI stories tend to over-explain and simplify plots, while human stories explore moral ambiguity and temporal complexity more. - Specific LLMs, including Claude, GPT, and Gemini, exhibit unique narrative fingerprints identified by the system.
- StoryScope analyzes discourse-level narrative features to identify AI-generated fiction.
- The pipeline achieved high accuracy in distinguishing human from AI stories and attributing to specific LLMs.
- AI stories show common patterns like over-explanation and simple plots, contrasting with human narrative diversity.
- Individual LLM models exhibit distinct narrative characteristics or 'fingerprints'.