Why do AI systems struggle to craft truly profound literary works? A fresh study conducted by researchers at the University of North Carolina at Chapel Hill sheds light on this issue. The core weakness of generative models, they argue, is that they strip characters of their enigmatic quality—AI-driven heroes lack mystery. By employing a novel assessment framework known as CASPER, the experts examined thousands of texts and demonstrated that neural networks tend to oversimplify psychological profiles in favor of a predictable outcome.
Why fiction remains a challenge for AI
Deep textual analysis. To analyze the texts, the researchers built the CASPER framework, which evaluates characters based on established literary principles. The system checks characters for realism, tracks their development as the plot progresses, and examines the final resolution and closure of conflicts. A survey of professional writers who use AI reveals that authors delegate routine tasks to the technology—such as searching for sources—but always craft the core traits of their characters themselves.
Artificial intelligence is making deeper inroads into publishing and the film industry, aiding authors in drafting novels or scripts. However, as a detailed literary analysis reveals, AI-generated characters come across as flat and predictable.
AI agents rely on familiar patterns and aim to neatly tie up all narrative threads, leaving no room for the reader’s imagination. In contrast, human writers effortlessly create contradictory characters whose motives remain unsolved even after the story ends.
Overview of the CASPER structure, including the sequence for building the corpus (top row), conducting experiments (middle row), and analyzing categories (bottom row).
It is precisely this sense of incompleteness and ambiguity that makes human literature memorable, compelling readers to reflect on the plot long after finishing the book. Even a massive increase in computational power and the size of language models fails to resolve this issue. The problem lies not in the volume of data but in the very understanding of what storytelling truly entails. The study has been published in the proceedings of a conference.
About the Author
