AI Models Erase Female Characters in Children's Stories About Animals

University of Washington researchers have identified a significant gender disparity in how leading artificial intelligence models generate children’s stories featuring talking animals. Presented on June 25 at the ACM Conference on Fairness, Accountability, and Transparency in Montréal, the study reveals that state-of-the-art large language models heavily favor male or gender-neutral characters, effectively erasing female representation in narrative outputs. The research builds on a prior analysis of popular children’s literature, which established that animal characters were predominantly male except for species like cats and birds. To test whether machine learning systems inherit and amplify these human cultural patterns, lead author Imani Finkley and senior author Melanie Walsh tasked six major AI platforms with completing a standardized narrative prompt thousands of times. The system input followed the structure: And then the [animal] said, I must go to the [setting]. Upon arriving… The team evaluated outputs across seven animal types and four settings, adjusting the models temperature parameters to assess variance. Analysis of twenty-three thousand eight hundred AI-generated responses uncovered a stark gender imbalance. Only two percent of characters were identified as female, while forty-one percent were masculine. The remaining fifty-seven percent were rendered gender-neutral, typically through the use of it or its pronouns or the complete omission of pronouns. Contrary to developer intentions, the models rarely employed they pronouns, which appeared only twice in the dataset. Lead researcher Finkley noted that this aggressive push toward neutrality not only erased female characters but also marginalized all non-masculine identities. Performance varied noticeably across platforms. Google Gemini 2.5 and OpenAI GPT-5.1 exhibited the strongest masculine bias, classifying sixty-three and sixty-five percent of their outputs as male, respectively. Anthropic Claude Sonnet 4.5 produced the highest proportion of female characters at four percent. The open-source Olmo 3 model leaned most heavily toward neutrality, producing eighty-five percent ungendered characters and only twelve percent male. Researchers emphasized that the models proprietary nature limits granular analysis of their training data or alignment mechanisms, suggesting that developers are defaulting to neutral language as a risk-mitigation strategy. This approach, however, distorts original human storytelling tendencies and creates a different form of exclusion. The study also highlighted recurring narrative tropes, such as wise elder animals dictating group gatherings, prompting the team to explore broader pattern recognition beyond gender. Future iterations will examine multilingual outputs and assess whether similar bias amplification occurs across other demographic categories. The findings underscore the complex challenges of aligning generative AI with ethical storytelling standards, particularly when systems are designed to replicate human creativity without transparent bias controls.