Exploring More-than-human Ethics of Large Language Models in Social Service Delivery

Large Language Models (LLMs) are increasingly used in social work practice, yet their ethical implications remain underexplored. This study addresses the gap by systematically reviewing literature on LLM-based tools in social service delivery. The review aims to identify the benefits, risks, and mitigation measures related to LLM use. Following PRISMA guidelines and employing a PICOC framework, we reviewed 34 studies published between 2019 and 2024 across multiple databases, including Web of Science, EBSCO, and Scopus. The results show that LLMs enhance efficiency in management, streamline frontline tasks, and support client needs. However, issues such as bias, lack of contextualization and explainability, and over-reliance pose ethical risks for both clients and professionals. We argue that Actor-Network Theory (ANT) helps reframe AI responsibility as distributed across human and non-human actors. It offers analytical tools to trace non-human agency and address ethical risks, highlighting the need for technical, organizational, and regulatory improvements.