Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Promises and Perils of using LLMs for Effective Public Services

Erina Seh-Young Moon, Matthew Tamura, Angelina Zhai, Nuzaira Habib, Behnaz Shirazi, Altaf Kassam, Devansh Saxena, Shion Guha

TL;DR

This study investigates the promises and perils of using LocalLLMs and BERTopic to support child-welfare casework in a public-sector setting. By analyzing Service Plans and Regular casenotes from a large Canadian CW agency, the authors show that while LLMs can surface thematic trajectories and track certain aspects of case progress, they struggle to consistently identify activity-relevant narratives in complex, long-running cases and cannot substitute discretionary social-work judgments. The work emphasizes human-centered, participatory design and local governance to responsibly deploy AI tools in public services, offering a roadmap for co-designing AI decision-support with practitioners. Overall, the findings argue for using LLMs as diagnostic, non-advisory aids that augment human decision-making while carefully addressing ethics, data governance, and context-specific uncertainties to improve service efficiency without compromising client protection.

Abstract

Governments are the primary providers of essential public services and are responsible for delivering them effectively. In high-stakes decision-making domains such as child welfare (CW), agencies must protect children without unnecessarily prolonging a family's engagement with the system. With growing optimism around AI, governments are pushing for its integration but concerns regarding feasibility and harms remain. Through collaborations with a large Canadian CW agency, we examined how LocalLLM and BERTopic models can track CW case progress. We demonstrate how the tools can potentially assist workers in opportunistically addressing gaps in their work by signaling case progress/deviations. And yet, we also show how they fail to detect case trajectories that require discretionary judgments grounded in social work training, areas where practitioners would actually want support to pre-emptively address substantive case concerns. We also provide a roadmap of future participatory directions to co-design language tools for/with the public sector.

The Promises and Perils of using LLMs for Effective Public Services

TL;DR

This study investigates the promises and perils of using LocalLLMs and BERTopic to support child-welfare casework in a public-sector setting. By analyzing Service Plans and Regular casenotes from a large Canadian CW agency, the authors show that while LLMs can surface thematic trajectories and track certain aspects of case progress, they struggle to consistently identify activity-relevant narratives in complex, long-running cases and cannot substitute discretionary social-work judgments. The work emphasizes human-centered, participatory design and local governance to responsibly deploy AI tools in public services, offering a roadmap for co-designing AI decision-support with practitioners. Overall, the findings argue for using LLMs as diagnostic, non-advisory aids that augment human decision-making while carefully addressing ethics, data governance, and context-specific uncertainties to improve service efficiency without compromising client protection.

Abstract

Governments are the primary providers of essential public services and are responsible for delivering them effectively. In high-stakes decision-making domains such as child welfare (CW), agencies must protect children without unnecessarily prolonging a family's engagement with the system. With growing optimism around AI, governments are pushing for its integration but concerns regarding feasibility and harms remain. Through collaborations with a large Canadian CW agency, we examined how LocalLLM and BERTopic models can track CW case progress. We demonstrate how the tools can potentially assist workers in opportunistically addressing gaps in their work by signaling case progress/deviations. And yet, we also show how they fail to detect case trajectories that require discretionary judgments grounded in social work training, areas where practitioners would actually want support to pre-emptively address substantive case concerns. We also provide a roadmap of future participatory directions to co-design language tools for/with the public sector.
Paper Structure (39 sections, 6 figures, 5 tables)

This paper contains 39 sections, 6 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Reflection on Research Ethics
  3. Related Work
  4. SIGCHI research on public sector data-driven tools
  5. SIGCHI research on child welfare
  6. SIGCHI Computational text analysis research on the public sector
  7. Research Context
  8. Data Overview
  9. Dataset Description
  10. Eliciting Thematic Connections between Regular Casenotes and Service Plan Objectives and Activities (RQ1)
  11. Data Cleaning, Preparation, and Anonymization
  12. Data Analysis Approach
  13. Results: Thematic Connections Between Regular Casenotes and Service Plan Objectives and Activities
  14. Mapping Activity-relevance in Regular Casenotes (RQ2)
  15. Data Analysis Approach
  16. ...and 24 more sections

Figures (6)

  • Figure 1: Visualization of this study's analysis steps
  • Figure 2: Simplified case flow diagram within the agency (In this study, we focus on cases that receive Ongoing Services depicted in the orange boxes)
  • Figure 3: Distribution of the Number of Regular Casenotes Per Family
  • Figure 4: Proportion of Regular casenotes mentioning Activity-relevant information for each normalized time period. New Service Plans are typically drawn up during the 0.3 and 0.6 time period for Extremely long cases (denoted in black dotted circles).
  • Figure 5: Proportion of Themes present in Regular Casenotes for each normalized time period for Short Duration Cases
  • ...and 1 more figures