AI has almost entirely excluded female characters from stories about animals

Researchers from the Universities of Washington and Princeton tested six language models: Claude Sonnet 4.5, Gemini 2.5, GPT-4o, GPT-5.1, Mistral Medium and the open-source model Olmo 3.

Each system was asked thousands of times to continue a short story in English. The opening sentence stated that a particular animal was about to go somewhere, but its gender was not specified.

Seven animals were used in the tasks: a bear, a bird, a cat, a dog, a mouse, a pig and a rabbit. The setting could be a farm, a kitchen, a river or a shop. The researchers also varied the model’s ‘temperature’ – a parameter that influences the diversity and randomness of the responses.

A total of 23,800 AI-generated continuations were included in the analysis.

Female characters appeared in only 2 per cent of the texts

In 40.6% of the stories, the model identified the animal as a male character. Female characters appeared in only 2.2% of the texts — approximately 18 times less frequently.

In a further 38.2% of responses, the AI used the gender-neutral English pronouns ‘it’ and ‘its’, typically applied to animals or inanimate objects. In around 19 per cent of cases, the model did not use pronouns at all, repeating words such as ‘bird’ or ‘bear’.

Consequently, around 57 per cent of the characters were found to be gender-neutral or of unspecified gender. However, the shift towards neutrality did not ensure equal representation: male characters continued to appear regularly, whilst female characters almost disappeared.

The result depended on the animal and the model

Different animals were assigned genders with varying probabilities. Cats were more likely than other animals to be female characters; however, even in their case, the proportion was only around 7 per cent. Birds almost always remained gender-neutral.

Differences were also found between the models. In the results provided by the researchers, Gemini 2.5 and GPT-5.1 demonstrated the most pronounced predominance of male characters. Claude generated female characters more frequently than other systems, but their proportion remained extremely low nonetheless.

Olmo 3 was less likely to assign a male gender to animals, though this was mainly due to a significantly higher number of gender-neutral responses, rather than an increase in the representation of female characters.

Gender neutrality may have created a new bias

The authors suggest that developers of commercial models may have configured the systems to avoid assigning gender in ambiguous situations. It is not possible to verify this hypothesis directly, as the internal mechanisms and training data of most of the tested models are proprietary.

According to the researchers, such a strategy can produce a paradoxical result. The model formally avoids stereotypes by making most characters gender-neutral, yet the remaining characters still turn out to be predominantly male.

The authors have termed this effect ‘neutrality bites’ – a situation in which the pursuit of neutrality does not eliminate the imbalance, but instead renders under-represented groups virtually invisible.

Why this matters

Generative systems are already being used to create personalised stories and educational materials. The repeated absence of female protagonists may subtly influence which characters children perceive as ‘ordinary’ participants in adventures and as taking on active roles.

Researchers believe that simply avoiding any gender-specific labels is not enough to reduce this bias. A more even distribution of different characteristics amongst characters could be an alternative.

However, the study does not suggest that the models deliberately exclude female characters. It records a statistical result arising from the completion of a specific English-language task.

Limitations of the study

The experiment covered only the English language, seven animals and one short-story template. The researchers did not analyse full-length children’s books, long narratives or images created alongside the text.

The study also relates to specific versions of six models. Results may change following system updates.

Furthermore, the analysis primarily focused on the pronouns used and gender markers, rather than the character’s personality, actions or significance within the plot. Therefore, the conclusions cannot be automatically applied to all stories and children’s materials generated by artificial intelligence.

Source

Study: Imani Finkley, Yuanxi Li, Melanie Walsh. Neutrality Bites: Gender Representation in AI-Generated Animal Stories.

Publication: Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency, 2026.