Proceedings of the Queer in AI Workshop

A Pranav, Alissa Valentine, Shaily Bhatt, Yanan Long, Arjun Subramonian, Amanda Bertsch, Anne Lauscher, Ankush GuptaNAACL 2025

TL;DR

This paper presents the proceedings of the Queer in AI 2025 workshop, highlighting research at the intersection of AI, NLP, and LGBTQIA+ perspectives. It addresses the critical need for inclusive AI development by showcasing work on identifying and mitigating queer bias in language models, understanding LGBTQ+ representation in media, and exploring the nuances of pronoun acquisition.

What Problem Does This Paper Address?

The rapid advancement of AI and NLP technologies, while promising, also carries risks for marginalized communities. When AI systems are developed without the input of those they impact, they can perpetuate existing biases and even create new harms. This is particularly relevant for the queer community, where AI systems might misrepresent, misinterpret, or even actively harm LGBTQIA+ individuals. The Queer in AI workshop aims to bridge this gap by providing a dedicated space for researchers, practitioners, and advocates to discuss and present work that centers queer and gender-diverse perspectives in AI and NLP. This includes addressing issues like the identification of hate speech targeting queer individuals, the detection of bias in large language models (LLMs) concerning LGBTQIA+ topics, and understanding the social and linguistic aspects of pronoun usage.

Methods

The workshop proceedings feature a variety of research methodologies. Several papers utilize LLM-based topic modeling to analyze representations of the LGBTQ+ community in media, such as reality TV shows. Other research focuses on evaluating the safety and ethical implications of LLMs, specifically examining their responses to LGBTQ+ language and their potential for dehumanizing queer groups in simulated interactions. Additionally, one paper employs a Bayesian approach to model the acquisition of pronouns and neopronouns, offering a computational perspective on linguistic development within the queer community. The methods range from established NLP techniques to novel evaluations of LLM behavior and computational linguistic modeling.

Datasets

The specific datasets used for training and evaluation are not explicitly detailed in the introductory sections of the proceedings. However, the papers themselves likely utilize a range of datasets relevant to their respective topics. These could include:

  • Media datasets: For analyzing representation, such as transcripts or content from TV shows like "RuPaul's Drag Race."
  • LLM interaction logs: For evaluating model responses to LGBTQ+ language and identifying harmful outputs.
  • Textual data: For hate speech detection and bias identification, potentially including social media posts, online forums, or curated datasets of LGBTQIA+-related content.
  • Linguistic corpora: For the study of pronoun acquisition, which might involve child language acquisition data or specialized corpora focusing on pronoun usage.

The proceedings do not appear to introduce new benchmarks, but rather leverage existing research and evaluation frameworks within the NLP and AI fields.

Key Insights and Ideas

  • LLMs can perpetuate harmful stereotypes: Research indicates that LLMs, without careful guidance, can generate responses that are not only unhelpful but also potentially harmful or dehumanizing towards LGBTQ+ individuals.
  • Representation matters: Analyzing media through an AI lens reveals how LGBTQ+ communities are portrayed, highlighting the need for more nuanced and accurate representations.
  • Understanding pronoun evolution is crucial: The study of pronoun and neopronoun acquisition sheds light on the dynamic nature of language and the importance of respecting evolving gender identities.
  • Dedicated spaces are vital: The existence and success of the Queer in AI workshop underscore the ongoing need for inclusive platforms that foster research and community building for underrepresented groups in AI.

Key Results

  • Topic modeling reveals representation patterns: Studies using LLM-based topic modeling can identify specific themes and stereotypes associated with the LGBTQ+ community in media, offering insights into how these groups are perceived and portrayed.
  • LLMs exhibit varied responses to LGBTQ+ language: Evaluations show that LLMs can struggle with understanding and responding appropriately to LGBTQ+ specific language, sometimes defaulting to unhelpful or biased outputs.
  • Bayesian models can capture pronoun acquisition: Computational models can effectively simulate and explain the process of acquiring pronouns and neopronouns, providing a framework for understanding linguistic development in this area.

Limitations and Future Work

The proceedings, being a collection of workshop papers, do not present a single, unified set of limitations. However, common themes and future directions implied across the research include:

  • Bias Mitigation: A significant area for future work is developing more robust methods to detect and mitigate biases against LGBTQ+ individuals in LLMs and other AI systems.
  • Data Scarcity: For certain niche areas within LGBTQIA+ research in AI, there may be a lack of large, diverse, and ethically sourced datasets, necessitating the creation of new resources.
  • Ethical Considerations: Continued focus on the ethical implications of AI research involving marginalized communities is paramount, ensuring that research benefits rather than harms these groups.
  • Interdisciplinary Collaboration: Future work will likely benefit from increased collaboration between AI researchers, linguists, sociologists, and LGBTQIA+ advocates to ensure comprehensive and impactful research.
  • Broader Impact: Expanding the reach and impact of Queer in AI research beyond academic conferences to influence industry practices and policy is a key future goal.

Summary updated

BibTeX
@inproceedings{pranav2025proceedings,
  author = {A Pranav and Alissa Valentine and Shaily Bhatt and Yanan Long and Arjun Subramonian and Amanda Bertsch and Anne Lauscher and Ankush Gupta},
  title = {{Proceedings of the Queer in AI Workshop}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure