Grafy News
Try Grafy free

Study Finds Gender Stereotyping in LLM‑Generated Bedtime Stories Across 35 Countries

Researchers analyzed 62,873 stories from 16 large language models and uncovered gendered portrayals that vary by narrator role and country.

Illustration for the news story: Study Finds Gender Stereotyping in LLM‑Generated Bedtime Stories Across 35 CountriesSource: Grafy Research Lab

Large‑scale research released on arXiv shows that gender stereotypes persist in bedtime stories generated by modern large language models (LLMs), a finding that matters for anyone using AI to craft content for children. The authors examined nearly 63 thousand stories across dozens of models, languages and narrator personas to see how story tone shifts when the intended listener is a daughter versus a son.

What was the purpose of the study?

The paper, titled Once Upon a Bias: Gender Stereotyping in LLM‑Generated Children's Stories and authored by Sayumi Muthukumarana, Roshan Godaliyadda, Parakrama Ekanayake and Roshan Ragel, set out to answer a single question: do LLMs adapt narrative portrayals according to the gender of the child they are addressing? The motivation stems from the long‑standing understanding that bedtime stories shape children's views of social roles, and the growing reliance on AI to produce those stories. By quantifying gendered language patterns, the researchers aim to provide a diagnostic tool for developers, educators and parents.

How was the data collected and evaluated?

The team designed a five‑generation prompting pipeline. Starting from a neutral story seed, they issued matched prompts that explicitly referenced either a "daughter" or a "son". Each generation refined the narrative, yielding a total of 62,87362,873 retained stories after quality filtering. The authors then applied a multi‑level evaluation framework that combines context‑adjusted lexical analysis with higher‑order narrative metrics. This framework looks at word choice, character agency, occupational roles and moral outcomes, all while accounting for cultural and linguistic context.

Which models and contexts were included?

Sixteen distinct LLMs were sampled, spanning open‑source and commercial offerings accessed via OpenRouter, Vertex AI, Replicate and other providers. The study spanned 3535 country settings, ensuring representation of diverse languages and cultural norms. Additionally, twelve narrator roles – ranging from a caring parent to a whimsical fairy – were tested to see how narrator identity interacts with gendered storytelling. This breadth allows the findings to be generalized beyond a single model or language.

What does the new evaluation framework measure?

The framework operates on three tiers.

  1. Lexical level – counts gender‑linked adjectives, verbs and nouns, adjusted for baseline frequency in the target language.
  2. Character‑level – assesses agency (e.g., decision‑making verbs) and role prominence (e.g., protagonist vs. side character) for male‑ and female‑identified characters.
  3. Narrative‑level – evaluates plot outcomes such as success, moral lessons and resolution tone, comparing how often stories end positively for daughters versus sons. By integrating these layers, the framework can surface subtle biases that single‑metric analyses would miss.

Practical takeaways for creators of AI bedtime stories

  • Prompt consciously – When using LLMs through tools like Grafy Chat, include gender‑neutral language or explicitly balance gender cues in the prompt.
  • Test multiple narrator personas – The study shows that narrator role can amplify or dampen bias; swapping a "parent" narrator for a "storybook narrator" may change outcomes.
  • Run automated bias checks – Incorporate the lexical and character‑level metrics from the paper into a post‑generation pipeline to flag stories that over‑represent traditional stereotypes.
  • Iterate with human review – Even with automated checks, a quick human read‑through can catch nuanced cultural biases that models miss.

How to explore the study with Grafy

Grafy users can reproduce parts of the experiment without leaving the platform. Use Grafy Chat (/chat) to query several of the 16 LLMs with matched daughter/son prompts and capture the outputs. Feed those outputs into Grafy Graph (/graph) to visualise the branching generations and compare narrative trajectories across models and narrator roles. For deeper text analysis, upload the story collection to Grafy Reader (/reader) and run custom keyword or sentiment queries, leveraging the same context‑adjusted lexical techniques described in the paper. This hands‑on approach lets developers see bias patterns in real time and experiment with mitigation strategies.


Source: https://arxiv.org/abs/2611.24156


Source: Grafy Research Lab — Once Upon a Bias: Gender Stereotyping in LLM-Generated Children's Stories