Large language models (LLMs) can generate written stories with unprecedented efficiency, but their outputs are highly dependent on their training data. This paper presents a research project in which we plan to analyze biases in AI-generated stories by examining how LLMs replicate societal stereotypes, especially regarding culture and gender inequalities. On a corpus of stories generated on diverse topics, characters, in different genres and languages, we will use content analysis, lexicography, and natural language processing tools, including explainability methods, to uncover and highlight biases in stories generated by LLMs. Findings will contribute to developing ethical AI systems that promote diverse and accurate representation in storytelling, helping to prevent the amplification of harmful stereotypes in AI-generated stories.
Escouflaire et al. (Wed,) studied this question.