The Automated Administrative State

Paper by Danielle Citron and Ryan Calo: “The administrative state has undergone radical change in recent decades. In the twentieth century, agencies in the United States generally relied on computers to assist human decision-makers. In the twenty-first century, computers are making agency decisions themselves. Automated systems are increasingly taking human beings out of the loop. Computers terminate Medicaid to cancer patients and deny food stamps to individuals. They identify parents believed to owe child support and initiate collection proceedings against them. Computers purge voters from the rolls and deem small businesses ineligible for federal contracts [1].

Automated systems built in the early 2000s eroded procedural safeguards at the heart of the administrative state. When government makes important decisions that affect our lives, liberty, and property, it owes us “due process”— understood as notice of, and a chance to object to, those decisions. Automated systems, however, frustrate these guarantees. Some systems like the “no-fly” list were designed and deployed in secret; others lacked record-keeping audit trails, making review of the law and facts supporting a system’s decisions impossible. Because programmers working at private contractors lacked training in the law, they distorted policy when translating it into code [2].

Some of us in the academy sounded the alarm as early as the 1990s, offering an array of mechanisms to ensure the accountability and transparency of automated administrative state [3]. Yet the same pathologies continue to plague government decision-making systems today. In some cases, these pathologies have deepened and extended. Agencies lean upon algorithms that turn our personal data into predictions, professing to reflect who we are and what we will do. The algorithms themselves increasingly rely upon techniques, such as deep learning, that are even less amenable to scrutiny than purely statistical models. Ideals of what the administrative law theorist Jerry Mashaw has called “bureaucratic justice” in the form of efficiency with a “human face” feel impossibly distant [4].

The trend toward more prevalent and less transparent automation in agency decision-making is deeply concerning. For a start, we have yet to address in any meaningful way the widening gap between the commitments of due process and the actual practices of contemporary agencies [5]. Nonetheless, agencies rush to automate (surely due to the influence and illusive promises of companies seeking lucrative contracts), trusting algorithms to tell us if criminals should receive probation, if public school teachers should be fired, or if severely disabled individuals should receive less than the maximum of state-funded nursing care [6]. Child welfare agencies conduct intrusive home inspections because some system, which no party to the interaction understands, has rated a poor mother as having a propensity for violence. The challenges of preserving due process in light of algorithmic decision-making is an area of renewed and active attention within academia, civil society, and even the courts [7].

Second, and routinely overlooked, we are applying the new affordances of artificial intelligence in precisely the wrong contexts…(More)”.