“I've never seen a glass ceiling better represented”: Bias and gendering in LLM-generated synthetic personas from a participatory design perspective

This study examines synthetic personas generated by Large Language Models (LLMs) and their implications, focusing on how these personas encode and perform gendering. Traditional personas carry implicit power and agency, making their accuracy and inclusivity essential. However, delegating persona creation to generative AI raises concerns about bias, representation, and ethical design. Poorly designed personas risk reinforcing stereotypes, marginalizing certain groups, and embedding biases into the design process. Using a mixed-method approach – combining direct inquiries with four LLMs and participatory workshops – we analyze gender bias in synthetic personas. Drawing from feminist theory, Human–Computer Interaction (HCI), and Participatory Design (PD), both societal, normative, and representational biases were identified. As a result of this, we argue that synthetic personas should not be used as direct stand-ins for real users but instead reframed as objects of critical inquiry. They can serve as provocations—tools that challenge assumptions and expose biases in LLM-generated outputs. Furthermore, this study underscores the need to move beyond exclusively expert-driven evaluations by incorporating user perspectives directly. By doing so, the evaluation process becomes richer, more representative, and better equipped to identify biases that might otherwise be overlooked.