UW researchers say AI is coming for kids’ books, but it’s biased
Russell Samora, 38, of western Massachusetts was reading a children’s book with his daughter when the thought crossed his mind: Why are the bears always boys?
He remembered thinking, “Wait a minute, there’s a bear and a rabbit in the story, but there’s absolutely no reason for the bear to be referred to as a boy.”
That’s when Samora started noticing a trend. Most of the animals in the books he read to his daughter were male. “It was something I couldn’t unsee,” he said.
He became curious why animal characters are usually boys. A data journalist himself, Samora teamed up with experts, including Melanie Walsh, an assistant professor of informatics at the University of Washington, to determine which genders were assigned most often to animal characters in children’s books.
Now, Walsh and doctoral students Imani Finkley and Yuanxi Li have taken those questions one step further by exploring what happens when artificial intelligence gets involved. Out of the 300 books Samora and his colleagues looked at in their original study published last year, researchers found he/him pronouns were used twice as often as she/her pronouns. Their research included 30 popular books from each decade since the 1950s, plus one category for books released before then.
They found certain animals were more likely to be described as male or female. While bears were mostly male, cats, on the other hand, were more likely to be female.
4 replies
AI will just take whatever bias is existing in the training data and, wanting to be plausible, will end up amplifying it. This process only compounds as AI trains on data generated by AI.
Related: Tell me about a popular female werewolf.
I see this bias when dicking around on a friend's home-built AI character generator, too. Simply by calling a character "a man" they become more dominant, take over conversations and are more likely to "correct" the user, while "women" are submissive and more likely to ask questions.
I'm not surprised AI has doubled down on all of our worst biases as a society. I do hope that something good comes from exposing it.
Non-paywall version