AI-Generated Synthetic Personas and Gender Representation in Large Language Models: Participatory and Feminist Perspectives on Emergent Technologies

Abstract

Large language models (LLMs) are increasingly being integrated into creative and decision-making processes, raising critical questions about their role in shaping knowledge and understanding. This dissertation examines the emerging phenomenon of LLM-generated synthetic personas (SP) and its implications for design processes. By delegating persona creation to generative AI, designers are confronted with critical questions about accuracy, inclusivity, and ethical representation. Poorly designed personas risk perpetuating stereotypes, marginalizing certain groups, and producing products that fail to address the needs of diverse user bases. This dissertation is animated by an overarching concern: What happens when the tools we use to represent users are themselves shaped by the biases we are trying to overcome, and how can participatory and feminist perspectives help us both expose and move beyond that tension? Bringing together four empirical studies and cross-study conceptual contributions, the present body of work focuses on what SP do: How they perform gender, how they can be repurposed as critical and speculative tools, and how participatory and feminist principles can both expose their limitations and redirect their use toward more inclusive ends. Combining participatory evaluations with statistical methods, I investigate how gendering is performed in LLM-generated SP, i.e. the ways in which these systems reproduce, reinforce, or obscure normative assumptions about gender. Drawing on Speculative and Critical Design, the work then moves to exploring how LLM-generated SP can function as speculative tools to challenge assumptions and norms in Human-Computer Interaction and equity-oriented research—from co-designed (queer) SP that surface gender-exclusive classifications in digital systems, to interactions with LLM-enacted SP that support broader inclusive technology evaluation and design. The findings reveal a central tension between the representational risks of LLM-generated SP and their generative potential when used interactively and critically. Across the studies, SP are reframed not as proxies for real users but as objects of critical inquiry; provocations that can surface assumptions and expose biases in both the models and the designers engaging with them. If generative AI is to play an expanded role within design, this dissertation argues it must be approached from a critical and feminist perspective, one that foregrounds pluralism, situatedness, and power. A central conceptual contribution is, thus, a bidirectional model of PD-with/for-AI, reconceptualizing the relationship between Participatory Design (PD) and Generative AI as a reflexive and recursive loop. SP may offer value and increase representation when used reflexively, but their uncritical adoption risks reproducing dominant narratives and undermining the democratic commitments of participatory and inclusive design.