TLDR: A qualitative study interviewed 13 Los Angeles policymakers in homeless services to gauge their openness to using AI for matching individuals to housing resources. Findings indicate general openness to AI if thoughtfully designed and used with human oversight, despite concerns about data quality, implementation complexities, and operationalizing transparency.
A new qualitative study delves into the perspectives of policymakers in Los Angeles regarding the integration of artificial intelligence (AI) into homeless services, specifically for matching individuals to scarce housing resources. The research, titled “Toward AI Matching Policies in Homeless Services: A Qualitative Study with Policymakers,” explores the potential gains and drawbacks of such systems, focusing on efficiency, fairness, and transparency.
The study highlights the pressing issue of homelessness in Los Angeles, where over 75,000 people were experiencing unsheltered homelessness as of January 2024, while only a fraction of permanent housing units exist. Currently, Los Angeles utilizes the Vulnerability Index – Service Prioritization Decision Assistance Tool (VI-SPDAT) to prioritize individuals. However, this tool has faced scrutiny due to evidence of racial bias and the fact that it was not originally designed for resource matching.
With the recent advancements in AI, researchers have proposed data-driven algorithms to improve resource allocation in homelessness services, potentially addressing issues like racial bias. However, the reception and adoption of these algorithms by practitioners, and their real-world consequences, remain unclear. This study aimed to bridge that gap by interviewing 13 policymakers who directly influence housing allocation policies in LA.
Policymakers’ Openness and Concerns
The study found that, even with an awareness of various complicating factors, policymakers generally welcome the idea of an AI matching tool, provided it is thoughtfully designed and used in conjunction with human decision-makers. Before the interviews, many policymakers had mixed sentiments about AI for matching, but after discussions, their perceptions either remained positive or became more positive.
One of the key benefits identified by policymakers is the potential for AI to reduce the tedious workload on overburdened caseworkers, thereby improving efficiency and making parts of the housing process simpler. They also expressed hope that AI could help reduce human bias in the matching process, while still emphasizing the importance of human involvement in what is fundamentally a human service.
However, the policymakers also voiced significant concerns and identified several complexities. The real-world housing system in LA is intricate, involving factors like specific eligibility requirements for housing units, the need for responsive caseworkers, and the unique needs of individual clients. Many of these nuances, such as a client’s exact disability or the specific stage of their housing journey, are not always captured in existing data systems like the Homeless Management Information System (HMIS).
External forces also play a role, including geographic boundaries for resource access, restrictions for undocumented individuals, and even direct political influence on housing decisions. These factors highlight the limitations of a purely algorithmic approach without human oversight.
Data Quality: A Foundational Challenge
A major concern raised by policymakers was the quality and trustworthiness of the HMIS data itself. Descriptions of the data ranged from “horrible” to “really crappy,” citing human error, lack of standardized data practices across agencies, and insufficient follow-up on client outcomes. There’s a significant worry that any AI system trained on such data could perpetuate existing biases, including the racial biases already observed in the VI-SPDAT and in who receives services through street outreach programs.
Policymakers noted a cultural resistance to data practices within social work, where the focus is often on direct human interaction rather than meticulous data entry. This lack of standardized practices could lead to disparities in data quality, potentially disadvantaging clients from agencies with poorer data habits.
Despite these challenges, some policymakers envisioned AI tools that could specifically address data issues, such as making data more accessible for providers or populating information based on verifiable sources to improve accuracy.
Fairness and Transparency: Divergent Views
When presented with a case study of an AI matching policy (from Tang et al., 2023) that aimed to maximize successful exits from homelessness while ensuring fairness, policymakers appreciated the potential equity gains. However, there was no consensus on which fairness metric an AI system should prioritize: “allocation parity” (equal rates of receiving housing resources across protected groups) or “outcome parity” (equal rates of successfully exiting homelessness across protected groups).
Similarly, opinions were divided on the proposed transparency features of the AI system, such as clients knowing their place in a queue or estimated wait times. While some saw value in empowering clients with this information, others worried about managing expectations and the potential for disappointment if timelines shifted. The complexity of the housing system, unlike a simple deli queue, makes such transparency challenging to operationalize effectively.
Also Read:
- Assessing AI’s Readiness for Critical Social Decisions: A Look at Homelessness Resource Allocation
- EMPATHIA: A New AI Framework for Dignified Refugee Integration
Moving Forward with AI in Homeless Services
The study concludes that there is a clear willingness among Los Angeles policymakers to explore AI tools for improving homeless services. However, for successful adoption, AI researchers and developers must collaborate closely with practitioners to address critical issues. This includes improving data quality and standardizing data practices, building trust through continuous evaluation and auditing of AI system performance, and engaging multiple stakeholders (including clients) to define the ultimate desired properties of an AI matching system.
The research emphasizes that AI tools should be designed to assist human decision-makers, not replace them. Future work needs to explore how human decision-making and AI can best integrate to achieve optimal efficiency, equity, and transparency gains. This study provides valuable insights for technologists aiming to build responsible algorithmic systems to support decision-making in low-resource scenarios like homeless services. For more details, you can refer to the full research paper here.
