Understanding Public Agencies' Expectations and Realities of AI-Driven Chatbots for Public Health Monitoring

Research Background and Issues

• Identified Problems or Challenges: Public health monitoring has traditionally relied on manual methods such as phone calls or home visits. However, due to resource constraints, this approach struggles to achieve frequent and widespread coverage. Additionally, hardware-dependent technologies (e.g., sensors), while helpful in expanding monitoring capabilities, often impose new labor burdens and maintenance tasks on staff with differing priorities. AI-driven chatbots, such as CareCall, have been proposed to address these issues, but there is limited research on their actual impact and staff perceptions.
• Significance: Against the backdrop of an aging population and rising instances of solitary deaths, public health monitoring is critical for supporting vulnerable groups in society. AI technology has the potential to significantly improve monitoring efficiency and coverage. If effectively integrated, it could reduce reliance on human labor.
• Research Motivation and Related Work: Current studies primarily focus on the needs of frontline workers regarding AI technologies, with limited attention to the perspectives of decision-makers and administrative staff. This research aims to analyze the practical application of AI-driven tools (e.g., CareCall) to explore the experiences and needs of different stakeholders, thereby gaining a more comprehensive understanding of the actual effects of AI technologies in public services and the hidden labor costs involved.

Solution

• Proposed Method or Solution: This study conducts a case analysis of the AI chatbot CareCall and performs in-depth interviews with 21 public institution staff members to understand the discrepancies between its expected functionalities and real-world applications. CareCall, powered by a large language model (LLM), simulates the tone of a social worker to conduct health check-ins with users.
• Innovations:
  1. Investigates how AI chatbots can be deployed in public health monitoring and evaluates their real-world impact on multiple stakeholders.
  2. Introduces the theory of "articulation work" to identify the human labor demands and shifts in new labor tasks associated with CareCall's implementation.
  3. Highlights the potential of LLM-driven open-ended dialogue systems in uncovering unmet care needs.
• Implementation Steps and Key Technologies:
  1. Collect and analyze data on CareCall deployment and document its specific impacts on decision-making, monitoring, and administrative tasks through interviews.
  2. Evaluate CareCall's effectiveness in expanding monitoring coverage and frequency while reducing hardware maintenance requirements.
  3. Utilize LLMs (e.g., HyperCLOVA) with long-term memory capabilities to understand users' health conditions through continuous conversations and generate personalized inquiries.

Research Outcomes

• Specific Findings:
  1. Expanded Coverage: CareCall significantly increased the reach of public health monitoring through telephone networks, particularly in low-resource settings, outperforming hardware-dependent technologies.
  2. Discovery of Unexpected Care Needs: CareCall's open-ended dialogue enabled the identification of previously overlooked social service or health needs, such as mental health issues or post-disaster recovery requirements.
  3. New Labor Burdens: CareCall introduced new labor demands, including handling interruptions in user engagement (e.g., missed calls) and managing the complexity of coordinating follow-up work for frontline staff.
  4. Limitations: While the technology expanded monitoring coverage, mismatched resources increased the burden on frontline workers. Additionally, discrepancies between the chatbot's capabilities and its limitations raised issues in managing public and staff expectations.
• Advantages Over Existing Solutions:
  • Avoided the high costs and maintenance burdens of hardware devices, leveraging existing telephone networks for large-scale deployment.
  • Offered personalized and open-ended conversation features, which outperformed traditional rule-based chatbots in building trust with users and uncovering unexpected information.
• Experimental or Evaluation Results:
  • Monitoring frequency for CareCall users increased from a few times per month to weekly, with a significant rise in the number of individuals covered.
  • The cost of monitoring using CareCall was substantially lower than hardware-dependent solutions (e.g., motion sensors or smart plugs).
• Limitations and Future Directions:
  • Limitations: Insufficient human resource planning; over-reliance on assumptions about users' willingness to interact with AI; task allocation often led to imbalanced workloads among staff.
  • Future Directions:
    1. Develop standardized guidelines to assist decision-makers in evaluating and planning AI system implementations.
    2. Explore how passive sensing (e.g., environmental data) and chatbots can complement each other to optimize health monitoring.
    3. Establish more robust user fallback mechanisms to address interruptions in usage.
    4. Enhance transparency in communicating the actual requirements and potential labor burdens of AI systems.

Conclusion

This paper highlights the potential and challenges of AI-driven chatbots in public health monitoring. The study reveals new labor demands arising from real-world deployments and offers specific recommendations for expanding AI applications in social services. Its unique value lies in analyzing the impact of AI technologies from the perspectives of multiple stakeholders and providing insights into optimizing technology design and deployment planning.