AI children's books show extreme gender bias, UW study finds

Artificial intelligence models used to generate children's stories are amplifying gender bias in unexpected ways, according to new research from the University of Washington.

When UW researchers prompted six major AI models—including ChatGPT, Claude, and Gemini—to complete children's stories featuring animal characters, the systems assigned male pronouns and names 41% of the time but female identifiers in just 2% of cases. The remaining 57% avoided gender markers entirely, using neutral terms like "it" or repeating the animal's name.

The disparity means AI-generated stories feature male animal characters roughly 20 times more often than female ones when gender is specified at all.

Why it matters

As AI-generated children's books flood self-publishing platforms like Amazon and tech companies roll out story-generation features, these biases directly shape what young readers encounter. Research shows children who don't see themselves represented in stories internalize the message that they don't matter—making algorithmic gender erasure a developmental concern, not just a technical curiosity.

Building on human bias research

The AI study extends earlier work by data journalist Russell Samora and UW assistant professor of informatics Melanie Walsh examining gender patterns in traditional children's books. That research analyzed 300 popular titles spanning seven decades and found male pronouns appeared twice as often as female ones.

When the team surveyed 1,300 people using story prompts featuring bears, birds, cats, dogs, mice, pigs, and rabbits, respondents favored male pronouns for all seven animals—revealing deep-seated human biases that AI systems appear to learn and distort.

The gender erasure problem

While corporate AI developers may be attempting to sidestep bias by avoiding gendered language, Walsh and doctoral students Imani Finkley and Yuanxi Li argue this approach backfires.

"They've basically erased female animal characters," Walsh said in an August university release. "So they're not only amplifying our human biases, but they're twisting them in strange, unexpected ways."

The opacity of commercial models makes it difficult to determine whether gender-neutral defaults stem from deliberate design choices or emergent behavior in training data. Walsh noted that closed systems like ChatGPT can't be scrutinized with the same rigor as open-source alternatives.

Real-world implications

Michelle Martin, the Beverly Cleary professor for children and youth services at UW's Information School, emphasized that representation matters particularly in early childhood materials. "If you don't see it in what you read, you get the idea that you don't count," she said.

Some institutions are responding to AI content concerns. Seattle Public Library has committed to avoiding AI-generated materials in its collection, though managing librarian Kate Sellers acknowledged detection remains challenging, especially for digital formats that vendors don't consistently label.

The findings were first reported by The Seattle Times, with the underlying research building on work published in The Pudding last year.