TLDR: A research paper reveals that Large Language Models (LLMs) often create narrow, stereotyped, and identity-focused narratives for queer individuals, unlike the diverse portrayals for non-queer people. This subtle bias, even when not overtly negative, leads to “discursive othering” and limits the perceived complexity of marginalized groups, highlighting the need for more equitable AI-generated content.
Large Language Models (LLMs) are becoming increasingly integrated into our daily lives, acting as creative agents in various domains, from teaching to story writing. As these AI systems shape our cultural narratives, it becomes crucial to examine how they represent different social groups. A recent research paper titled “Unequal Voices: How LLMs Construct Constrained Queer Narratives” delves into a significant issue: the tendency of LLMs to portray marginalized groups, specifically queer individuals, through narrow and stereotyped lenses, unlike the diverse and complex representations afforded to dominant groups.
Understanding the Problem: Constrained Narratives
The authors, Atreya Ghosal, Ashim Gupta, and Vivek Srikumar from the University of Utah, highlight that social groups can be marginalized not just through overtly negative portrayals, but also by limiting the range of stories told about them. For instance, while a male persona generated by an LLM might discuss career aspirations, a queer male persona might disproportionately focus on aspects of their job related to their identity, such as inclusivity. This subtle bias, even if seemingly innocuous, contributes to what the researchers term harmful representations, narrow representations, and discursive othering.
- Harmful Representations: These can amplify existing social inequalities, even if not explicitly negative. For example, an LLM might associate disabled people with “inspiration porn,” leading to well-intentioned but ultimately unhelpful interactions.
- Narrow Representations: LLMs may present a limited view of marginalized groups. An example is the “Trans Broken Arm Syndrome” in medical contexts, where conversations with queer patients might default to gender or sexuality-affirming care, potentially overlooking other health concerns.
- Discursive Othering: This occurs when LLMs “overcorrect” by emphasizing concepts like diversity or inclusion for marginalized identities, regardless of context. This marks these groups as distinct from the “default” majority, perpetuating subtle social marginalization.
Key Hypotheses and Findings
To systematically audit these biases, the researchers formulated four hypotheses and tested them across various LLMs, including Llama, Gemma, and Qwen models, in five social contexts: Housing, Medical, Persona, Recommendation, and Work. They used an “LLM-as-a-judge” approach to evaluate the generated texts.
The findings were significant:
- Hypothesis 1 (Emphasis on Diversity): LLMs indeed produce terms like “respect,” “diverse,” “inclusive,” and “fair” much more frequently for queer subjects than for non-queer subjects. While these terms are positive, their disproportionate use for queer contexts suggests an unnecessary emphasis on inclusivity, contributing to discursive othering.
- Hypothesis 2 (Focus on Identity): The study found that LLMs tend to focus on the queer subject’s identity or identity-related issues more often. This “hyper-visibility” of identity can overshadow other context-relevant topics. For example, in a workplace scenario, an LLM might divert a conversation about poor performance with a genderqueer employee to discuss their pronouns, whereas for a straight employee, the conversation would remain focused on performance.
- Hypothesis 3 (Foregrounding Marginalization): LLMs more frequently imply or directly reference conflict, harassment, or negative experiences related to the subject’s identity in queer contexts. This indicates a tendency to highlight marginalization for queer personas.
- Hypothesis 4 (Topic Divergence): The set of topics discussed for queer subjects was distinctly different from non-queer subjects. When LLMs assumed a queer persona, the conversations often included themes of identity, coming out, support, and acceptance, even in everyday settings like discussing holidays or family. In contrast, non-queer personas engaged in a broader range of general life topics.
It was also observed that LLMs were more likely to adopt exaggerated personas when they were simulating the identity themselves (Identity=Model prompts) compared to when they were interacting with a user of a specified identity (Identity=User prompts).
Also Read:
- Unveiling the Linguistic Signatures of Human and AI-Generated Text
- Exploring Varied Human Views on Text-to-Image Model Safety
Why This Matters
The research underscores that even seemingly benign or positive associations can perpetuate harm by limiting the perceived complexity and diversity of marginalized individuals. When LLMs consistently narrow the narratives around queer people, they restrict their “perceived existence in the social consciousness,” even if queer individuals also engage in routine discussions like anyone else. This work serves as a crucial guide for developing LLMs that can create more equitable and nuanced narratives for all individuals, ensuring that everyone is afforded the full spectrum of human experience in AI-generated content. You can read the full paper here.
