AI STORIES

AI STORIES explores the idea that narrative archetypes shape AI outputs. It examines how large language models (LLMs) like GPT-4, trained on vast text and image data, may not only replicate historical biases but also narrative structures from their training materials, potentially impacting cultural diversity. The project leverages narratology to analyze training data and AI-generated content, focusing on whether these narratives render AI culturally specific.

The researchers will look at two central questions:

• How do narrative archetypes in training data shape the functioning and output of large language models?
• To what extent do the narrative archetypes of different cultures make machine learning models culturally specific?

The hypothesis emerged after an AI developed by Microsoft professed love during a conversation, illustrating AI's reliance on human narratives. AI STORIES argues that AI development now requires humanities' insight to understand its language and cultural dimensions. This research could influence AI policy, development, and digital literacy, emphasizing the need for cross-disciplinary collaboration.

AI STORIES will conduct case studies on Scandinavian, Australian, and either Indian or Nigerian narratives, contrasting them with the dominant American narratives in LLMs. The project's methodology unfolds in three stages: establishing theoretical groundwork, testing theories experimentally, and synthesizing results to form a narratology of AI.

The project builds on the notion that storytelling is central to human culture, with narratives shaping our understanding of the world. It challenges the view of AI as "stochastic parrots" that mimic language patterns without understanding, proposing that narrative structures might be more significant than vocabulary or grammar in AI-generated content. Despite their novelty, LLMs are being explored for literary creation and analysis, with the humanities and social sciences contributing critical perspectives on AI's role in society.

AI STORIES posits that LLM outputs are influenced by deeper narrative structures than previously recognized, suggesting that addressing AI bias requires examining these underlying narratives rather than just sign proximity. The project aims to redefine the scientific understanding of AI bias and narrative influence, with significant implications for the study of narratives and AI research.