Neutrality Bites: Gender Representation in LLM-Generated Animal Stories
Imani Finkley, Yuanxi Li, and Melanie Walsh
In ACM Conference on Fairness, Accountability, and Transparency (FAccT), 2026
(Accepted)
We investigate gender bias in large language model-generated narratives by examining how six leading LLMs assign gender to anthropomorphic animal characters across 23,800 stories. While models frequently use gender-neutral language (38.2% on average) or omit gender entirely (19% on average), significant masculine bias emerges when gender is assigned: female characters appear in only 2.2% of stories compared to 40.6% featuring male characters. We argue that prioritizing neutrality as a bias-mitigation strategy may inadvertently erase marginalized identities, and advocate instead for methods that distribute social possibilities more equitably across narrative subjects.