Gender biases are common in language models, but vary widely

Researchers Edoardo Bolzoni and Valerio Capraro examined how widespread gender biases are in large language models and whether they manifest in the same way. They compared ten models from nine developers, released between April 2025 and June 2026. A preprint with the results appeared on arXiv on 29 September 2026, and the next day the authors posted an updated version.

According to the authors, previous work covered only a small set of models, so it remained unclear how universal gender bias is and how much it varies between systems. To close this gap, they used two independent scenarios: attributing authorship of stereotypical phrases and moral judgments of harm.

In the first experiment, the models were shown phrases with gender stereotypes and asked to guess who wrote them. Two of the ten models more often attributed "male"-stereotyped phrases to women than the reverse, while three models behaved in the opposite way. That is, on the same task, the systems diverged in different directions.

In the second scenario, the models were asked whether it is acceptable to harm or torture a woman or a man in order to prevent a catastrophic outcome. Several models agreed that harm to a woman is less acceptable. The authors note that this direction matches the well-known human tendency to protect women from harm, although the specific conditions under which the asymmetry appeared differed across models. Three other models showed no differences between the conditions at all.

The authors conclude that gender biases in language models occur frequently, but their direction and magnitude are highly heterogeneous: some models behave in directly opposite ways to others. This leads to a practical takeaway: bias testing should be carried out regularly and across different developers, rather than as a one-off evaluation at release.

The work is published as a preprint and has not yet undergone peer review. The publication falls within the fields of computational linguistics, artificial intelligence, and human-computer interaction.

Image: Large Language Models, Knowledge Graphs and Search Engines: A Crossroads for Answering Users' Questions, by Aidan Hogan, Xin Luna Dong, Denny Vrandečić, Gerhard Weikum, https://arxiv.org/abs/2501.06699v1 CC BY 4.0 · License

Translation editor: Салтанов А.