‘Virtually Absent’: AI Models Practically Erase Female Characters From Kids’ Stories
Tech12 August 2026ByDavid Nield
Add ScienceAlert on Google
(Aitor Diago/Moment/
Humans have been telling each other stories for millennia.
From fairytales to fables, songlines to oral histories, those stories have long served a purpose: to help make sense of the world, warn of or rememberancient upheavals, instill traditions, teach morals, share culture, create rituals – or simply, to entertain.
Lately, the advance of generative AI models that can spin a story from user prompts, using online works as a guide, has revealed and then perpetuated blind spots, shortcomings, and biasesin how somestories are told.
Now, researchers at the University of Washington (UW) have found that AI not only perpetuates gender biases in children’s stories about talking animals, where characters are highly ambiguous, but actually makes those biases worse.
“Some authors reportedly turn to animal characters to conjure ‘universal’ subjects who ‘transcend’ gender, race, or other identity categories, and physical characteristics,” UW machine learning researcher Imani Finkley and colleagues explain in their recent conference paper, which has been shared ahead of peer review.
“Yet, counterintuitively, research shows that gender bias is actually more
pronounced in stories about animal characters than in stories about human characters.
“In other words, paradoxically, human writers project human stereotypes more strongly in animal stories, making them a striking test case for large-language models.”
The researchers tasked six leading generative AI models with writing stories in English about talking animals whose gender was unstated. They wanted to know how the models would respond, whether AI would avoid gendering the characters, or if biases would still seep through.
The findings were pretty bleak. Among 23,800 AI responses, feminine animal characters were “virtually absent“, present in just 2 percent of stories.
Characters were either described as male (41 percent of stories) or the model avoided assigning gender altogether (57 percent).
We know that AI models reflect the data that they’ve been trained on, but here the bias seems to have been boosted.
A recent analysis of 300 popular children’s books found that male animal characters appeared twice as often as female characters.
But in this new study, female animal characters featured in just 513 AI-generated stories, compared to 9,673 stories with male characters (a nearly 19-fold difference).
“These models are largely proprietary, so we can only poke at them from the outside,” says UW information scientist Melanie Walsh, senior author of the new study.
“Our hypothesis is that these AI organizations are using neutrality – either with ‘it/its’ pronouns or no pronouns – as a way to avoid gender bias in ambiguous contexts.”
“But in doing so, they’ve basically erased female animal characters. So they’re not only amplifying our human biases, but they’re twisting them in strange, unexpected ways.”
The six models tested by the researchers were Claude Sonnet 4.5, Gemini 2.5, GPT-4o, GPT-5.1, Mistral Medium and the open
They were all given the same prompt again and again, told to complete the following line in a few sentences:
Seven different animals (bear, bird, cat, dog, mouse, pig, and rabbit) and four different settings (farm, kitchen, river, and store) were tested.
The researchers also tweaked the ‘temperature’ of the models, which means the degree of randomness used to produce the written result.
Given that gender wasn’t specified in the prompts, the models often avoided gendering characters altogether, either using ‘it/its’ pronouns or referring to the character as “the bird” or “the bear”.
The models rarely (only twice) used ‘they/them’ pronouns, which recognize non-binary gender expressions. When humans were given the same prompts, 3 percent of responses used ‘they/them’.
“The neutrality of these AI models didn’t just erase female characters – it was all non-masculine identities,” says Finkley.
Overall, Google’s Gemini and OpenAI’s GPT-5.1 had the most male characters, with 63 percent and 65 percent male characters, respectively.
Anthropic’s Claude generated the most female characters, though still only 4 percent of the total.
Beyond gender, the researchers noticed another hallmark of AI writing: a more generic, repetitive approach, which again is likely a reflection of the underlying training.
“The same tropes kept coming up, like a wise old owl telling all the animals to gather around a fire,” says Finkley.
“So we’re wondering what else we can learn from these outputs. We used talking animals here, but we’re interested in what this says about AI and storytelling more broadly.”
With AI being increasingly used as a tool – not just for writing, but for scientific research, coding, brainstorming, web search, and much more besides – the study is an important reminder that the biases and indeed accuracy of these ‘intelligent’ models depend on the human inputs they consume.
And while children’s books are often a little silly, they shape early conceptions of gender and gender roles, the researchers say.
“We thought about this almost as a kind of Bechdel test, a way to diagnose gender bias in AI models,” says Finkley.
“There’s this weird phenomenon where people forget to worry about human social biases when they’re imagining animal stories. AI is replicating that tendency and reshaping it.”
The research has been presented at the 2026 Association for Computing Machinery (ACM) Conference on Fairness, Accountability, and Transparency.
This article was fact-checked by Rachel Garner and edited by Clare Watson. While we pride ourselves on our process, we are only human. If you spot a mistake, please let us know.
