Understanding Frontline Workers’ and Unhoused Individuals’ Perspectives on AI Used in Homeless Services

Title of the Paper

Understanding Frontline Workers’ and Unhoused Individuals’ Perspectives on AI Used in Homeless Services

Paper Information

• Research Domain: Human-Computer Interaction, Applications of Artificial Intelligence in Public Services, Social Computing
• Keywords: Homeless services, AI decision support systems, participatory design, public algorithms, risk assessment

Research Background and Problem

• Identified Problems or Challenges: The study reveals that AI decision support systems (ADS) are being rapidly deployed in homeless services to prioritize the allocation of scarce housing resources. However, key stakeholders related to these systems, particularly the unhoused individuals and frontline workers directly affected, have had minimal opportunities to participate in the design process or provide feedback. Existing designs may fail to fully reflect community needs and could even lead to negative consequences.

• Significance: Globally, homelessness affects over 1.8 billion people, including significant populations in developed countries. The use of AI in homeless services has the potential to exacerbate existing social issues, particularly harming already marginalized groups. Therefore, broad and effective stakeholder engagement is critically important.

• Research Motivation and Related Work: While existing research has extensively explored the ethics, bias, and algorithmic transparency of AI in the public sector, there is insufficient research on the rapid expansion of AI systems in homeless services. This study aims to better understand the needs and concerns of direct stakeholders through a novel participatory design approach.

Proposed Solution

• Proposed Solution: The authors introduced a participatory feedback and design method called "AI Lifecycle Comicboarding." This approach is an adaptation of the "comicboarding design method," specifically tailored to help groups with low literacy or limited technical knowledge understand complex AI system designs.

• Innovative Contributions:

  1. Integrating comicboarding with various stages of the AI system lifecycle to enable participants to provide feedback on specific design components (e.g., problem definition, data selection, model design).
  2. Offering a structured process for gathering feedback at different design stages, using visualization and simplified technical reporting to make AI systems comprehensible to non-technical groups.

• Implementation Steps and Key Techniques:

  1. Preparing Comicboarding Content: Create comicboards illustrating the main stages of the AI system lifecycle, including problem formulation, data selection, model definition, and deployment.
  2. Engaging Participants: Conduct one-on-one research interviews with unhoused individuals and frontline workers, using the comicboards to discuss design ideas and gather feedback.
  3. Cross-Referencing Opinions: Share anonymized feedback from different stakeholder groups during interviews to facilitate cross-group discussions.
  4. Result Analysis: Apply Reflexive Thematic Analysis to code and synthesize themes from collected audio and textual data.

Research Outcomes

• Specific Outcomes:

  1. The comicboarding method elicited extensive and detailed feedback on AI systems, including aspects such as model selection, goal setting, and data quality.
  2. Established frontline workers and unhoused individuals as critical sources of feedback, even in the absence of technical expertise.
  3. Interactions with stakeholders uncovered hidden assumptions and decisions embedded in algorithmic design.

• Advantages Over Existing Solutions:

  1. The method overcame barriers related to technical understanding, enabling non-technical users to seamlessly participate in AI system design feedback.
  2. Provided stakeholders with a direct avenue to influence research and potential social policies, enhancing decision-making transparency and community inclusivity.

• Experimental or Evaluation Results: A total of 1,023 codes were collected during interviews, generating rich themes such as the misalignment between design goals and the needs of unhoused individuals, and systemic gaps or biases in the data relied upon by algorithms.

• Limitations and Future Directions: Limitations:

  • The method is primarily suited for late-stage design feedback, and its application during early design stages remains unexplored.
  • A small number of participants may not fully represent diverse communities.

  Future Directions:

  • Extend the method to other public service domains (e.g., child welfare or mental health services).
  • Develop toolkits to facilitate broader adoption of the method.
  • Investigate more effective ways to directly integrate stakeholder feedback into AI design processes.