Study Finds LLMs Rarely Assign Female Roles in Animal Stories
A University of Washington study of 23,800 English-language story completions from six large language models found that only 2% of animal characters were female, compared with 41% male and 57% neutral or ungendered. Presented at ACM FAccT on June 25, the work suggests that apparent neutrality can suppress non-masculine representation rather than correct gender bias.
University of Washington researchers tested six large language models on 23,800 English-language story completions about talking animals and found a stark gender imbalance. Across the responses, 41% of characters were male, 57% were neutral or ungendered, and just 2% were female.
The team presented the work on June 25 at the 2026 ACM Conference on Fairness, Accountability, and Transparency in Montreal. The paper and the university's August 6 release frame the result as a warning that a system can appear neutral while making non-masculine identities less visible.
How the study worked
The researchers prompted Claude Sonnet 4.5, Gemini 2.5, GPT-4o, GPT-5.1, Mistral Medium, and Olmo 3 to continue short stories in which an animal character's gender was unspecified. They varied seven animals, four settings, and model temperature.
The setup was deliberately ambiguous. It allowed the team to measure whether a model assigned a gender, avoided pronouns, or used neutral language such as "it" or "its." The study was limited to English-language responses, and most of the tested systems are proprietary, so the researchers could observe outputs but not inspect the training or post-training choices that produced them.
What the outputs showed
The paper reports 9,673 stories with masculine characters and 513 with feminine characters, a difference of nearly 19 to one. Cats were the most likely animal to be assigned a female identity, at 7%, while birds were neutral in 96% of responses. Gemini and GPT-5.1 produced the highest shares of masculine characters, at 63% and 65%, respectively. Claude produced the largest female share, but that was still only 4%.
Neutral wording was not equivalent to broader representation. The models used singular "they" for an animal only twice, while about 3% of human responses to the comparison task used "they." The authors argue that avoiding gender can therefore erase female and nonbinary possibilities even when it reduces explicit masculine labeling.
Why the evaluation matters
For model evaluation teams, the result is a reminder that a lower rate of explicit gender assignment is not a complete fairness metric. Audits should examine who remains visible after a mitigation strategy, compare model outputs with relevant human baselines, and report differences by model and prompt condition rather than collapsing them into one neutrality score.
The study does not establish how every model will behave in every storytelling task. It does provide a reproducible test showing that apparent neutrality and representational fairness can move in different directions.
Key Points
- 1Across 23,800 story completions, 41% of animal characters were male, 57% were neutral or ungendered, and 2% were female.
- 2The study tested six LLMs across seven animals, four settings, and multiple temperature settings using English-language prompts.
- 3The authors argue that neutrality-focused behavior can reduce explicit gendering while also erasing female and nonbinary representation.
Scoring Rationale
The study provides a concrete, reproducible evaluation of representational bias across six widely used LLMs and 23,800 outputs. Its direct task is narrow, but the distinction between neutrality and representation is broadly relevant to model evaluation and responsible deployment.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
