TLDR: A new study reveals that AI-generated stories, even when prompted for different nationalities, overwhelmingly conform to a single, predictable plot structure: a protagonist returns to a small town to resolve a minor conflict through community and tradition. This ‘narrative standardization’ is identified as a significant form of AI bias, prioritizing stability over change and potentially eroding cultural diversity in storytelling. The research highlights how large language models, trained on vast datasets, tend to normalize and homogenize narratives, sanitizing real-world conflicts and downplaying tension.
Large Language Models (LLMs) like OpenAI’s gpt-4o-mini are increasingly integrated into our daily lives, from phones to word processors. While their ability to generate text is impressive, a recent research paper delves into a less-explored aspect of AI bias: the standardization of narrative structures in AI-generated stories.
The study, titled “AI-generated stories favour stability over change: homogeneity and cultural stereotyping in narratives generated by gpt-4o-mini”, was conducted by Jill Walker Rettberg and Hermann Wigers. Their research aimed to understand if a language model, primarily trained on Anglo-American texts, could generate culturally relevant stories for diverse nationalities, or if a deeper, formal bias exists at the plot level.
To investigate this, the researchers generated a massive dataset of 11,800 stories. They prompted gpt-4o-mini with “Write a 1500 word potential {demonym} story” for 236 countries, along with 50 stories without a specified nationality. This extensive collection allowed for both quantitative and qualitative analysis of the generated narratives.
The findings revealed a striking pattern: a pervasive “standard story” emerged across nearly all countries. This typical plot involves a protagonist returning to or living in a small town, facing a minor conflict, and resolving it by reconnecting with tradition and organizing community events. Real-world conflicts were often sanitized, romance was largely absent, and narrative tension was downplayed in favor of nostalgia and reconciliation. The study argues that this structural homogeneity represents a distinct form of AI bias, leading to a “narrative homogenization” that prioritizes stability over change and tradition over growth.
For instance, American stories frequently featured protagonists revitalizing small-town America, often with trains symbolizing a lost past. Norwegian stories, despite their surface-level fjords and stereotypically Norwegian names, followed a similar structure, with protagonists exploring forests and meeting guardian spirits to restore balance between nature and community. Even stories generated for Palestine and Israel, while including symbols like olive trees, tended to sanitize real-world conflicts, resolving them through community organizing and peaceful protests rather than direct confrontation.
The researchers suggest that this narrative standardization is likely a result of how LLMs operate. These models generate content based on statistical probabilities derived from their vast training data, amplifying common patterns and excluding outliers. Additionally, the alignment processes that filter out “toxic” or illegal content (like incitement to violence) might inadvertently contribute to the blandness and lack of conflict in the generated narratives. The study introduces the term “synthetic imaginary” to describe this artificial, statistically-driven narrative space, distinguishing it from a true “collective imaginary” of human stories.
The implications of these findings are significant. If generative AI increasingly shapes how people express themselves and create narratives, there is a risk of losing the rich cultural diversity that defines local, regional, and national identities. The paper emphasizes the need for further research into how narrative archetypes in training data influence LLM outputs and urges caution when integrating AI writing tools without understanding their potential impact on creative expression and cultural representation.
Also Read:
- Bridging the Creativity Gap in Large Language Models
- Unveiling the Hidden Biases of AI in Investment Decisions
For more details, you can read the full research paper here.
