Large language models for social governance: An analytical framework and research agenda

Abstract

As social governance becomes increasingly complex, traditional governance approaches face challenges such as information asymmetry, fragmented coordination, delayed risk response, and uneven public service delivery. Large language models (LLMs) offer new possibilities for knowledge processing, decision support, and human–machine interaction in this context. This article provides a comprehensive and framework-oriented review of research on LLMs, digital governance, public governance, and computational social science. It develops a governance-oriented analytical framework that links LLMs capabilities, governance functions, operating mechanisms, governance outcomes, and risk constraints. The review examines representative application scenarios, including emergency management, information propagation, intelligent rule of law, data governance, public security, healthcare, and e-government. The analysis highlights uneven evidence maturity across these scenarios: bounded tasks such as policy retrieval, public consultation, information summarization, and administrative document processing have clearer deployment pathways, whereas high-stakes simulation, legal reasoning, public security assessment, healthcare triage, and cross-departmental decision support require stronger validation and oversight. LLM-enabled social governance also remains constrained by hallucination, algorithmic bias, privacy leakage, weak explainability, resource inequality, accountability gaps, and institutional adaptation challenges. This article argues that LLMs should be understood as decision-support infrastructures embedded in accountable, auditable, and public value-oriented governance institutions.