AI Models Nearly Erase Female Characters When Writing Kids' Stories, Study Finds

AI Gender Bias Emerges in Children's Storytelling

When AI models write kids' stories about talking animals, female characters nearly vanish. A University of Washington study led by Melanie Walsh tested six leading AI models across 23,800 story completions and discovered that just 2% of AI-generated animal stories featured female characters1

1

2

2

. The research, presented on June 25 at the 2026 ACM Conference on Fairness, Accountability, and Transparency in Montréal, reveals how AI alignment guardrails designed to eliminate gender bias have backfired, creating an unexpected pattern of gender representation in AI3

3

.

Extreme Shift Toward Gender Neutrality

The University of Washington study found that AI models took a sharp turn into extreme gender neutrality when completing children's stories. Across all responses, 57% of characters were assigned neutral or ungendered pronouns like "it," while male characters made up 41%1

1

. This pattern emerged from testing GPT-5.1, Gemini 2.5, Claude Sonnet 4.5, Mistral Medium, GPT-4o, and the open-source Olmo 3 model developed by researchers at Seattle's Allen Institute for Artificial Intelligence and the UW3

3

. Walsh explained that AI organizations appear to be using neutrality as a strategy to avoid gender bias in AI models, but this approach has essentially erased female animal characters rather than balanced representation1

1

.

Stark Differences Between AI Models

The gap between individual models was significant. GPT-5.1 and Gemini 2.5 produced the most stories with masculine bias, generating male characters in 65% and 63% of responses respectively2

2

. Olmo 3 leaned heavily into neutral framing, with 85% of responses using ungendered characters and only 12% masculine characters1

1

. Claude Sonnet 4.5 generated the highest proportion of female characters among all models tested, though that figure still maxed out at just 4%3

3

. These findings matter because AI systems like Google's Gemini Storybook now allow parents and teachers to generate personalized AI-generated children's stories on demand3

3

.

How Human Biases Compare to AI Outputs

The research builds on Walsh's earlier work analyzing 300 popular children's picture books alongside journalists from The Pudding. That study revealed a distinct masculine bias in traditional publishing: out of 13 common animal tropes, most defaulted to male unless the character happened to be a cat, duck, or bird1

1

. When Walsh's team gave 1,300 human participants the same story completion prompts used in the AI study, human readers leaned even further into male pronouns for every single animal tested1

1

. However, about 3% of human responses used singular "they/them" pronouns, while AI models used this inclusive language only twice across thousands of story completions3

3

.

Animal-Specific Gender Patterns

Specific animals triggered distinct patterns in AI-generated animal stories. Cats were assigned female pronouns 7% of the time, the highest of any creature tested1

1

. Birds defaulted to neutral language in 96% of responses, while a frog or wolf had over a 90% chance of being gendered male in the original children's books analysis3

3

. The researchers tested seven different animals—bear, bird, cat, dog, mouse, pig, and rabbit—across four settings: farm, kitchen, river, and store3

3

. Temperature settings and location didn't greatly affect model outputs overall, but the choice of animal significantly influenced gender assignment patterns.

Implications for AI Development and Storytelling

Imani Finkley, a UW doctoral student in the Information School and lead author of the study, emphasized that the neutrality of AI models didn't just erase female characters—it eliminated all non-masculine identities1

1

. The researchers view this experiment as a diagnostic tool, similar to a Bechdel test for evaluating how AI models handle gender representation in storytelling3

3

. Finkley noted that people often forget to worry about human biases when imagining animal stories, and AI is replicating that tendency while reshaping it in unexpected ways1

1

. The team observed recurring character tropes throughout the generated text, such as a wise old owl telling animals to gather around a fire, suggesting deeper patterns worth exploring beyond gender bias in AI models3

3

.

What This Means for Parents and Educators

Anyone using tools like Google Home's Interactive Storytime feature to generate animal tales should take note of these findings2

2

. With female leads appearing in roughly one in every 46 AI-generated stories, parents and teachers cannot rely on AI models to naturally produce balanced gender representation2

2

. If you want your child's story to feature a female lead, you'll likely need to explicitly request it in your prompt. The research team plans to expand beyond English-language prompts in future work and analyze other narrative patterns that emerged from the study3

3

. These findings fit a broader pattern of bias in AI systems, echoing other research showing chatbot decisions shift based on traits like age, religion, and gender2

2

.