The team developed an AI-powered explanation feature that effectively translates complex, multi-program policy calculations into clear and accessible explanations, enabling users to explore "what-if" scenarios and understand key factors influencing benefit amounts and eligibility thresholds.
The team developed an AI solution to assist benefit navigators with in-the-moment program information, finding that while LLMs are useful for summarizing and interpreting text, they are not ideal for implementing strict formulas like benefit calculations, but can accelerate the eligibility process by leveraging their strengths in general tasks.
The team conducted experiments to determine whether clients would be responsive to proactive support offered by a chatbot, and identify the ideal timing of the intervention.
The Atlanta Fed’s CLIFF tools provide greater transparency to workers about potential public assistance losses when their earnings increase. We find three broad themes in organization-level implementation of the CLIFF tools: identifying the tar- get population of users; integrating the tool into existing operations; and integrating the tool into coaching sessions.
This article analyses ‘digital distortions’ in Rules as Code, which refer to disconnects between regulation and code that arise from interpretive choices in the encoding process.
This paper describes results from fieldwork conducted at a social services site where the workers evaluate citizens' applications for food and medical assistance submitted via an e-government system. These results suggest value tensions that result - not from different stakeholders with different values - but from differences among how stakeholders enact the same shared value in practice.
CHI '14: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems
This paper introduces a method for auditing benefits eligibility screening tools in four steps: 1) generate test households, 2) automatically populate screening questions with household information and retrieve determinations, 3) translate eligibility guidelines into computer code to generate ground truth determinations, and 4) identify conflicting determinations to detect errors.
