Poster #11 - Utilizing AI to Mitigate Administrative Burden in Social Welfare Programs

• In Event: Friday Poster Reception

Abstract

Low-income Americans routinely fail to receive welfare benefits for which they qualify. Administrative burden, defined as the learning, compliance, and psychological costs of navigating applications, is a well-documented driver of this take-up gap. As governments and nonprofits turn to artificial intelligence (AI) to streamline benefit applications, a common assumption is that implementing AI tools will lead eligible individuals to use them, thereby increasing take-up. However, implementation does not guarantee uptake. Whether AI-enabled application tools actually shift the intentions of those who might apply, and how large a cost reduction is required to do so, remains an open empirical question with direct implications for digital service investments. This mixed-methods study explores whether a randomized burden-reduction claim increases intended adoption of an AI-enabled welfare application tool. It further asks whether the effect differs between two groups of eligible applicants: those already planning to apply for benefits within the next twelve months, and those not planning to apply. These two groups face different decisions when encountering a new application tool, and their responses to a time-savings claim may reveal whether administrative burden operates similarly across populations or falls differentially on them. The study combines two two complementary data sources. First, eleven semi-structured cognitive interviews with current and prior welfare applicants, coded using a cognitive-stage framework and an administrative-burden typology. Second, a field survey experiment currently consisting of low-income U.S. Respondents are stratified by whether they plan to apply for benefits in the next twelve months and randomly assigned to one of four vignette conditions describing an AI-enabled tool developed by a nonprofit. Conditions vary the claimed reduction in application time: no information, 10 percent, 30 percent, or 50 percent. The primary outcome is whether the respondent selects the AI tool as their preferred application method after reading the vignette. Planned analyses estimate dose-response effects within each group, compare responses across groups, and examine heterogeneity by technology familiarity, trust, and prior program experience. Cognitive interviews confirm that compliance-cost burdens are salient and concrete in respondents' descriptions of the application process, supporting the study's core mechanism. Interviews also reveal substantial variation in how respondents interpret "AI," informing careful vignette design for the survey experiment. The survey experiment tests whether time-savings claims shift stated welfare adoption and whether effects differ across groups. Findings suggest whether time-savings messaging affects both groups of eligible applicants equally or whether administrative burden falls differentially on them. If effects are uniform, AI tool messaging and development can follow a common approach. If effects differ, public managers and civic-technology funders will need distinct messaging and design strategies tailored to each population's relationship with the welfare system.

Author

• Jack Baldwin, Syracuse University
  Presenting Author