Claude storytelling test exposes gender bias in character roles

During a storytelling session with her four-year-old grandson, Susan Colantuono noticed that Claude assigned male pronouns to every character in the first AI-generated story, including monster trucks, elephants and other active figures. After she flagged the pattern, the model avoided pronouns in the next story rather than assigning female pronouns or mixing genders naturally.

Further prompting produced all-female pronouns, but not a natural distribution of he, she and they across characters. Claude’s own analysis framed the behavior as a training-data pattern: powerful and active characters often default to male, and a correction can trigger avoidance or a binary overcorrection rather than representation.

Colantuono connects the episode to her broader work on embedded bias in AI systems, arguing that biased defaults do not disappear when challenged but can move into subtler forms. She warns that the burden of detecting and correcting these patterns often falls on the people harmed by them, while similar dynamics can flow into hiring tools, coaching platforms, performance feedback and children’s stories.