Modern government is built on the civil service. The twentieth century consensus, emerging out of the New Deal, was based on a social safety net administered by civil servants to ensure fair administration and due process. Yet today, decades of denying agencies the legal flexibility and technical resources to modernize have created a crisis that fuels political backlash. This crisis is now driving hasty attempts at technological fixes: the “Department of Government Efficiency” seeks to dismantle the civil service through mass layoffs and rapid AI deployment. At the same time, understaffed state agencies are experimenting with AI for benefits administration, with one claiming its system “taught itself all of the [state’s] eligibility rules” and eliminated a massive claims backlog in one month. Both approaches can rush toward AI solutions without adequate safeguards, presenting a false choice between dysfunctional status quo and reckless technological disruption, when thoughtful modernization could preserve administrative due process while meeting modern challenges.
This Article charts a path beyond this impasse by examining how administrative law should govern AI’s evolving role in benefits adjudication. We trace how merit staffing requirements, designed to
ensure procedural fairness and insulate decisions from politics, have also impeded responsible innovation. These regulations produce contradictory guidance that leaves states uncertain about AI’s legal boundaries, forcing them to choose between violating timeliness mandates or merit staffing rules.
The pandemic intensified this dilemma, spurring widespread state experimentation—from fraud detection to chatbots to “auto-adjudication.” While AI tools can drastically reduce backlogs, increase outreach, and improve efficiency, they also introduce risks of error, opacity, and bias. Through a simple yet novel audit of state systems, we expose vulnerabilities in these largely unexamined experiments. Yet we also show, through the IRS’s evaluation of its phone help line that integrated machine assistance, that rigorous assessment of human-AI systems is both feasible and essential.
We argue that the path forward to vindicating constitutional and administrative law values lies in evaluation. While implicit in Mathews v. Eldridge, the failure to make this explicit has led to neglect. Evaluation as due process is not merely a bureaucratic exercise but a legal imperative—a means of operationalizing due process at scale and with flexibility that preserves the dignity, fairness, and accountability of administrative systems.