Talk Title
Queer Voices for Gender Inclusion, Too: Queer Realities Meet Harm Evaluation Frameworks for Generative AI

Abstract
Transgender and non-binary (TGNB) individuals disproportionately experience discrimination and exclusion from daily life. Given the recent popularity and adoption of generative AI and large language models (LLMs), the potential to further marginalize this population only grows. Assessing these systems for gender-inclusivity requires understanding how TGNB identities uniquely interact with dominant gender norms; crafting harm evaluations for TGNB persons must be guided by TGNB voices, different from gender binary-centric perspectives. Harm measurement frameworks inherently require centering TGNB voices to help guide the alignment between gender-inclusive Natural Language Generation (NLG) and whom they are intended to serve. In this talk, I discuss grounding LLM harm evaluation in the TGNB community and existing interdisciplinary literature to assess how the social reality surrounding experienced marginalization of TGNB persons contributes to and persists within Open Language Generation (OLG). This social knowledge serves as a guide for evaluating popular LLMs (ChatGPT, GPT-2, GPT-Neo, and OPT) on two key aspects: (1) misgendering and (2) gender disclosure. We discover a dominance of binary gender norms in the LLMs and consider what this means for understanding harms to users who don’t ascribe to the gender binary. Across the talk, I contextualize findings in commonly-used robotic manipulation datasets like OpenAI’s CLIP and explore what these insights mean for responsible robotics.
Paper: https://dl.acm.org/doi/10.1145/3593013.3594078

Speaker Bio
Anaelia Ovalle (they/them) is an Afro-Carribean, queer, and non-binary PhD candidate in Computer Science at the University of California, Los Angeles. Advised by Prof. Kai-Wei Chang, Anaelia’s research centers the interfacing of ML design with algorithmic fairness objectives and the broader socio-technical milieu. With particular emphasis on historically marginalized communities, their research synergizes across both algorithmic fairness and critical social theory to guide approaches in mitigating AI-driven sociotechnical harms. Their work operates at 2 resolutions: (1) inclusive natural language processing and representation learning (e.g. how do we center trans communities in NLG, what does nonbinary exclusion mean and look like in a language context?) and (2) expanding AI ethics praxis (in both research and broader institutional mechanization) through intersectionality and participatory design. They have previously interned across several Responsible AI teams including Meta, Amazon Prime Video, and Amazon Alexa. Their research with Alexa was just accepted to ACM FAccT 2023. Prior to their PhD, Elia received a BS magna cum laude in Data Science from the University of San Francisco.

This talk is supported by an LGBTQ+ Project Grant from the Spectrum Center.