Representation of fine-grained queer identity terms

Determine how well German and multilingual language models represent fine-grained queer identity terms, with the aim of assessing whether these systems capture such identities accurately and inclusively.

Background

The paper evaluates anti-LGBTQ bias in German and multilingual LLMs using both a German translation of WinoQueer and a culturally grounded community-gathered dataset. The results show substantial variation in bias across identity categories, particularly for less frequent or more fine-grained terms such as genderfluid, genderqueer, agender, and demisexual. The authors therefore leave unresolved the broader question of whether models represent these identity terms adequately, rather than merely associating them with stereotypical or offensive language.

References

It also remains an open question how well models represent fine-grained queer identity terms and whether alternative evaluation designs beyond paired sentences would improve validity.

— Evaluating and Mitigating Anti-LGBTQ Biases in German and Multilingual Language Models  (2608.30884 - Morch et al., 31 Aug 2026) in Section Conclusion