Using Generative AI to Help States Avoid SNAP Payment Error Penalties

• In Event: Leveraging Data and Partnerships to Help States Navigate SNAP Policy Changes

Abstract

New federal legislation (H.R. 1, 2025) requires state SNAP agencies to reduce their payment error rates below 6 percent within the next few years. If states fail to meet this target, they will be required to bear a portion of SNAP benefit costs, which for some states with high error rates could exceed $1 billion. Faced with this fiscal risk, some states are considering changes to eligibility pathways, such as scaling back Broad-Based Categorical Eligibility, and — in extreme cases — have even raised the possibility of ending SNAP administration completely, increasing food insecurity among people in poverty.

State agencies are therefore exploring ways to reduce errors quickly, but traditional responses such as expanding caseworker review are costly and difficult to scale: most SNAP administrative budgets are already strained, it can take years for new staff to master complicated eligibility rules, and application review often depends on unstructured materials and rapidly changing federal guidance. These constraints create a strong need for tools that can help agencies review applications more consistently and efficiently.

This paper asks whether a carefully designed generative AI system can help states identify likely application errors before caseworkers conduct full review, while maintaining equitable performance across groups. Generative AI is especially promising in this context because it can reason over unstructured application materials, identify apparent conflicts with current eligibility criteria, and synthesize issues for human reviewers. Yet the use of generative AI in this setting raises concerns about hallucination, bias, and caseworker overreliance, making rigorous validation essential before deployment in real-world benefits administration.

We study a generative AI-powered approach that uses plain-language, state-specific policy rules as a shared reference for administrators and large language models. The rules are designed to be readable and editable by non-technical staff, creating a more transparent and auditable basis for AI-assisted review compared to traditional software approaches used in public sector settings. Following these guidelines, the system analyzes submitted materials, flags potential inconsistencies for human follow-up, and drafts clarifying questions for applicants before a staff member makes an official eligibility determination; it is not designed to automate final benefit decisions.

To evaluate the approach, we construct a simulated corpus of approximately 10,000 SNAP applications designed to reflect the distribution of payment error types observed in routine quality-control processes. We assess the system's ability to detect likely errors, characterize false positives and false negatives, and compare performance across racial and ethnic groups. Results indicate that the system identifies likely errors with high accuracy and limited disparities in flagging rates across groups. These findings suggest that generative AI can be designed to complement, rather than replace, human caseworkers in high-stakes public benefits administration.
The paper contributes to current efforts to help states respond to changing federal SNAP requirements while preserving administrative fairness and program access. More broadly, it illustrates how state agencies can pair policy guidance, administrative data, and human oversight to deploy generative AI responsibly in frontline service delivery.

Author

• Alex Chohlas-Wood, New York University
  Presenting Author