Document Type
Poster
Publication Date
Fall 2025
Abstract
Frontier large language models (LLMs) generate fluent outlines but often converge on repetitive structural templates. This exploratory study examined inter-model diversity in thriller novel outlines produced by five LLMs (ChatGPT-5(5.1), Grok-4, Gemini-2.5 Flash, DeepSeek-v3, Claude Sonnet-4.5) under baseline and enhanced prompting conditions (N=50 outlines). Blind AI evaluations revealed moderate inter-model diversity (baseline mean 64.45%; enhanced 58.89%), declining relative to intra-model baselines and between phases, particularly in Protagonist and Plot dimensions. Quantitative and qualitative analysis identified phase-specific convergence: a "forensic procedural" template (analytical female protagonists, corporate conspiracies, Quest/Monster arcs) under the baseline prompt, shifting to a "domestic psychological thriller" template (uniform female caregivers in their 30s, tragedy arcs, moral ambiguity) under enhanced constraints. While the enhanced prompt eliminated baseline clichés (e.g., forensic accountants, tidy resolutions), it introduced new uniformities, resulting in comparable or reduced perceived diversity. Outliers were disproportionately contributed by Gemini Flash 2.5 and DeepSeek-v3. These findings indicate that expert prompting redirects rather than expands narrative convergence, highlighting structural limitations in current LLMs and informing future efforts toward greater range in AI-assisted storytelling.
Recommended Citation
Hernandez Brito, Marisol, "Many Stories, One Shape: Narrative Convergence in AI-Generated Fiction" (2025). IPHS 484: Senior Seminar. Paper 42.
https://digital.kenyon.edu/dh_iphs_ss/42
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.