Introduction
The Proceedings of the Queer in AI Workshop, held on May 4, 2025, in conjunction with the Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), marks a transformative moment for the field of Natural Language Processing (NLP). Historically, the Queer in AI initiative has served as a vital affinity group for LGBTQIA+ scientists. However, the 2025 workshop represents a significant maturation of this community, transitioning into a formal archival venue dedicated to the intersection of gender-diverse identities and artificial intelligence.
This collection of research addresses a critical gap in the AI landscape: the historical exclusion of marginalized stakeholders from the development and evaluation of language technologies. By providing a peer-reviewed space for queer-centric research, the workshop aims to mitigate the erasure, bias, and active harm often perpetuated by systems built without the participation of the communities they impact.
Methodological Diversity in Queer NLP
The inaugural proceedings showcase a broad methodological toolkit, reflecting the multidisciplinary nature of Queer AI research. Rather than focusing on a single algorithmic approach, the contributors utilize diverse techniques to probe and improve modern AI systems:
- LLM-Based Topic Modeling: Researchers leveraged Large Language Models (LLMs) to analyze cultural representations, such as those found in media like RuPaul’s Drag Race, to understand how queer narratives are encoded in data.
- Behavioral Probing: Studies investigated the nuances of how LLMs respond to LGBTQ+ language and sexual interactions, identifying specific failure modes in model behavior.
- Classification and Detection: Advanced LLM architectures were employed to identify and categorize anti-LGBTQIA+ hate speech within user-generated content, aiming for more robust moderation tools.
- Computational Linguistics: Utilizing Bayesian frameworks, researchers modeled the acquisition and usage of neopronouns, providing a formal linguistic account of how language models and humans adapt to evolving gender-neutral terminology.
Specialized Datasets and Benchmarks
A primary contribution of this workshop is the introduction of specialized datasets designed to capture the nuances of queer identity—data often missing from standard NLP benchmarks. These include:
- Cultural Transcripts: Curated data from queer media used for representation and sentiment studies.
- Safety and Guardrail Benchmarks: Evaluations specifically designed to test the tension between "guardrails" (safety filters) and "guidance" (inclusive responses) in LLM interactions involving queer identity.
- Hate Speech Corpora: Specialized collections of text curated to detect subtle and overt anti-LGBTQIA+ sentiment.
- Linguistic Modeling Data: Datasets focused on the structural usage patterns of neopronouns to assist in more inclusive language modeling.
Key Insights: Participation as a Safety Requirement
The proceedings argue that the inclusion of queer researchers is not merely a matter of diversity metrics, but a fundamental requirement for AI safety. Without the active involvement of queer scientists in the evaluation process, the specific risks posed to the community—such as dehumanization in sexualized AI interactions or the erasure of non-binary identities—remain invisible to developers.
The transition to an archival format (featuring five archival and three non-archival papers) signifies the emergence of "Queer AI" as a rigorous academic sub-discipline. This shift allows for the formal dissemination of research that captures gender-diverse perspectives, which are frequently overlooked in mainstream AI evaluation frameworks.
Future Directions and Limitations
While the 2025 workshop is a milestone, the field remains in its early stages. The organizers highlight several areas for growth:
- Scaling the Discipline: With a modest number of archival submissions, there is a clear need for the field to expand to match the recent surge in interest regarding queer-centric NLP.
- Institutional Support: The workshop continues to rely heavily on the grassroots efforts of queer rights activists. Sustaining this momentum will require deeper institutional support from major AI and NLP organizations.
- Proactive Design: Future work must move beyond reactive bias detection. The goal is to integrate queer perspectives into the foundational study, development, and evaluation of NLP technologies, fostering a shift toward proactive, inclusive design.
By formalizing these contributions, the Queer in AI Workshop at NAACL 2025 establishes a foundation for a more equitable and representative future in artificial intelligence research.
Summary updated