Georgetown Law
Technology Review

The federal government is expanding its use of artificial intelligence (AI) to perform key human tasks. But the Administrative Procedure Act (APA), enacted in 1946, is ill-equipped to handle the use of AI by administrative agencies.1The Administrative Conference of the United States (ACUS) defines artificial intelligence as systems that “tend to have characteristics such as the ability to learn to solve complex problems, make predictions, or undertake tasks that heretofore have relied on human decision making or intervention.” Admin. Conf. of the U.S., Administrative Conference Statement #20: Agency Use of Artificial Intelligence, 86 Fed. Reg. 6616, 6616 n.1 (Jan. 22, 2021) [hereinafter Statement #20]. Given its dynamic nature, approval of AI software in the rulemaking process is unlikely.2See David Engstrom & Daniel Ho, Algorithmic Accountability in the Administrative State, 37 YALE J. REG. 800, 842 (2020) (“[E]ach model may be distinct so that the model reviewed at one stage (notice and comment) may already be substantively quite different when deployed . . .. These challenges become more acute as agencies adopt more advanced forms of machine learning that are dynamic in nature.”). At the same time, the AI black box problem means that courts are unable to evaluate the “reasoned decisionmaking” of an AI adjudicator in the same way that they currently review determinations by Administrative Law Judges (ALJ).3The black box problem “occurs whenever the reasons why an AI decision-maker has arrived at its decision are not currently understandable . . . because the system itself is not understandable.” Jordan Joseph Wadden, Defining the Undefinable: The Black Box Problem in Healthcare Artificial Intelligence, J. MED. ETHICS, at 764, 767 July 21, 2021. Despite these challenges, the federal government continues to encourage the implementation of AI technologies by administrative agencies. This places the federal judiciary on an AI collision course.