University of Washington researchers found that six leading AI models generated male characters in 41% of responses when asked to write stories about talking animals, while just 2% of characters were female. The majority - 57% - were gender neutral or ungendered, according to a study presented at the 2026 ACM Conference on Fairness, Accountability, and Transparency in Montréal.
The research builds on earlier work by UW assistant professor Melanie Walsh, who examined how 300 popular children's books gendered their animal characters. In that analysis, most animal characters were male - frogs and wolves came in at over 90% male - while cats, ducks, and birds trended slightly more female. Human participants who completed story prompts showed a similar masculine bias: every animal tested was more likely to be described as male.
For the new study, Walsh and her students gave six leading AI models the same prompt they had given human participants: "And then the [animal] said, 'I must go to the [setting].' Upon arriving…" The models completed the prompt thousands of times across seven animals and four settings.
Neutrality came at the cost of female characters
The gender-neutral responses were almost entirely either ungendered references like "the bird" or "it/it/its" pronouns. "They/them" pronouns appeared only twice in 23,800 responses to refer to a single animal character. In the human study, about 3% of responses used "they/them."
"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," said Walsh, the study's senior author. "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 gender breakdown varied by model: Gemini and GPT-5.1 had the strongest masculine bias at 63% and 65% of responses. Claude produced the most female characters, at 4%. Olmo 3, an open-source model from the Allen Institute for Artificial Intelligence and the UW, had the fewest masculine characters (12%) and the most neutral (85%).
Different animals, different outcomes
Adjusting the models' "temperature" - the randomness of generated text - and changing the story setting didn't significantly affect output. The animal did: cats were the most likely to be gendered female, at 7%, while birds were 96% neutral.
"These models are largely proprietary, so we can only poke at them from the outside," Walsh said, noting that the researchers couldn't inspect training data or internal decision-making directly.
Why this matters for writers and creatives
Parents and teachers can now generate personalized illustrated stories using tools like Google's Gemini Storybook, and the models trained on human writing inherit its patterns - including this new twist on an old bias. For writers and other creatives who use AI to generate drafts or brainstorm characters, this study offers a concrete lesson: default settings will silently shape character representation in the text.
The study is limited to English-language responses, and the team plans to explore other languages in future research. They've also noticed story templates that repeat and are worth examining. "The same tropes kept coming up, like a wise old owl telling all the reading animals to gather around a fire," said Claudia Finkley, a UW doctoral candidate in the Information School. "We thought about this almost as a kind of Bechdel test, a way to diagnose gender bias in AI models."
For professional creatives and writers who rely on generative tools, the findings show that AI's output isn't neutral - it's a dataset with its own biases that require active checking. Individual prompts or settings the story to avoid falling into the same patterns. If you're producing AI-assisted content, it's worth examining gender representation directly in the output, not assuming the default is balanced.
Your membership also unlocks:
