Simulated Stakeholders: Using Critical Frameworks to Examine LLM agents in early-stage design

There is a growing trend of using large language models (LLMs) to augment design activities such as needfinding. While LLM systems offer many potential benefits, using them to represent stakeholder voices also poses many risks. This paper borrows critical frameworks from other fields to examine Nurse Amaka, an LLM-based chatbot persona system developed to simulate engagement with stakeholders in humanitarian engineering contexts (i.e., with nurses in a rural Nigerian clinic). Using five critical frameworks: situated knowledges, sociotechnical imaginaries, delegated voice, epistemic injustice, and decolonial studies, we analyze the outputs of Nurse Amaka. The results highlight how uncritical deployment of LLMs for needfinding can inadvertently obscure and misinterpret stakeholder perspectives, reshape epistemic authority, and reproduce colonialist power imbalances. We conclude with Design Justice-based guiding questions and recommendations to help designers and stakeholders to navigate the risks of synthetic stakeholder representation and co-create tools and practices that take advantage of the benefits of LLMs while promoting stakeholder participation and voice.