As courts and administrative agencies encounter uncertainty posed by the post-Chevron era, a few foundational principles remain in place and serve as guides through doctrinal disruption. For one, courts have the authority and capacity to say what the law is. Likewise, if Congress unambiguously mandates a specific agency action, including the means to perform that act, jurists and scholars alike agree that Congress’s mandate must be followed. Absent that statutory clarity, the deference afforded to an agency’s interpretation of ambiguous language varies in light of several factors. Among legal scholars, the general thinking is that if Congress granted an agency the authority to determine the meaning of an ambiguous statute, then the judiciary should not infringe on Congress’s directive. Congress, though, rarely specifies the exact procedures agencies should use in interpreting and refining statutes. This ambiguity creates a contest for interpretative authority among courts and agencies. Whether agency reliance on artificial intelligence (AI) to inform their actions and interpretations should give them an upper hand in this skirmish has so far remained an open question.

　　Delay in answering this question is understandable and raising it now may seem to put the trailer before the truck. Agencies have yet to rely on AI to conduct the very important task of statutory and regulatory interpretation and drafting. My argument is that they eventually will. The combination of advances in AI capacities and AI adoption suggests that AI interpretations will become an agency norm sooner rather than later. Bruce Schneier and Nathan Sanders of Harvard contend that existing AI tools can already generate proposed regulatory changes, analyze the likely effects of that regulation, and develop strategies to increase the odds of its promulgation. Agency officials keen on accelerating their regulatory agendas would be hard-pressed not to turn to such useful tools.

　　There’s a case to be made that increased use of AI would generate “better” interpretations by agencies. Generative AI tools may lend agency interpretations a gloss of neutrality—a boon in a time of hyper-partisanship. On the other hand, reliance on AI tools may undermine the very purpose of delegation to experts who have acquired certain wisdom through experience administering the law. This essay argues the latter is the case.

　　The flaws associated with AI decision-making render it unworthy of any judicial deference when used by agencies to interpret statutes or regulations. AI interpretations detract from rather than augment any basis for respect for an agency’s understanding of a law or regulation. Before agencies test the limits of judicial tolerance for interpretations of the law predominantly informed by AI, it is important to establish a “Reality Doctrine”—AI interpretations merit no special weight and, if anything, should be subject to heightened judicial scrutiny.

　　Recent disruptions to the status quo in administrative law combined with ongoing advances in AI make now the right moment to address the idea of agency use of AI to interpret the law. The Supreme Court’s recent decision in Loper Bright Enterprises v. Raimondo effectively decided the interpretative war in favor of courts, while leaving some space for agencies to win individual interpretative battles via the power to persuade. Loper Bright was not revolutionary, so much as a rediscovery of a prior era of administrative law. Though the Court overruled some previously foundational aspects of administrative law, it left in place more established principles. More specifically, the Court recognized that the judiciary has long accorded due respect to Executive Branch interpretations that reflected the lessons gleaned from the “practice of the government.” In this new era, agencies have better odds of earning judicial respect if they adopt and adhere to processes that allow for well-reasoned judgments and incorporate documented findings from prior administration of the law. A new tool—generative AI—may aid in not only identifying those findings but also in steering agency decisions and interpretations. What persuasive power, then, should be afforded to an agency interpretation generated or informed by AI?

　　Judicial deference to agency actions and interpretations turns on a finite list of theories related to the processes used by that agency. Deference can also turn on whether the question is substantive or procedural. Courts tend to be more deferential to an agency’s choice as to how to pursue an end that is within their substantive authority. With few exceptions, the means, rather than the ends, determine the persuasive power of an agency’s interpretation. A flawed policy will receive the same deference as an ideal policy so long as the agency officials act on sound, deliberate procedures. This approach reflects but does not penalize the fact that bureaucrats are humans, not supercomputers.

　　Courts have tried to show deference to humanity in other ways, as well. Congressional intent, theoretically an aggregation of the intent of its very human members, is one theory of deference. Accountability to very human voters is yet another theory. Agency expertise, derived from the judgment, experience, and thoroughness of very human bureaucrats, is one more theory. The consistency and validity of decisions made by those bureaucrats may also affect the level of deference afforded to agency actions. What level of deference each of these theories receives and how courts rank these theories has changed over time, but this list has remained fairly fixed when referred to by scholars and relied on by jurists.

　　Unchecked encouragement of agencies using AI is misguided and irresponsible unless and until all legal actors understand how those uses fit into our broader systems of administration and judicial review.

　　Those theories do not explicitly cover how much persuasive power, if any, should be afforded to agency interpretations made or largely informed by AI. How those deference theories apply to AI interpretations has likewise been undertheorized and understudied. Much of the scholarship on this topic has focused on less substantive issues, such as discrimination that may occur as a result of government use of biased AI systems. Sooner than later, though, courts will have to fill in those legal lacunas.

　　Agencies have already started incorporating AI into more processes and decisions, but legal scholars have not given the issue enough thought. Even before OpenAI released ChatGPT-3, agencies extensively relied on AI. Little evidence suggests this trend will stop or even slow officials in other jurisdictions, including certain states, who have already started using AI to draft laws and author judicial opinions. What’s more, actors within the US government have called for increasingly substantive uses of AI by regulatory agencies. Some federal judges have already started to challenge opposition to the use of AI in drafting judicial decisions. The lid has come off AI’s Pandora’s box.

　　Pressure will continue to mount on agencies to join other public entities, foreign and domestic, in the incorporation of AI. These jurisdictions will pave the way in showing how AI can alleviate administrative burdens and further policy aims. They may also render the public more tolerant of an expanded role of AI in government work, as is being seen in some municipalities around the world. Congressional proposals and executive orders to ease and accelerate agency use of AI will likewise raise the odds of agencies exploring how to use AI. Shalanda Young, the director of the Office of Management and Budget, recently directed agencies to “increase their capacity to responsibly adopt AI, including generative AI, and take steps to enable sharing and reuse of AI models, code, and data.” .For now, though, current and planned uses of AI by agencies are fairly insignificant. OMB itself, for example, regards AI as “a helpful tool for modernizing agency operations and improving Federal Government service to the public.”

　　Rather than gamble that agencies will remain immune from pressure to use AI in more and more legally substantive fashions, the better bet is that agencies will become increasingly reliant on AI to take on even the most influential decisions, absent legal or policy reasons to avoid such uses of AI. Three key existing theories of deference apply to agency interpretations generated or substantively informed by AI, but based on the nature of AI, one must conclude that such interpretations hold no persuasive power.

　　The Reality Doctrine, on the other hand, recognizes that AI-generated interpretations lack all of the attributes that have traditionally justified judicial deference. In turn, the Doctrine demands de novo review of any AI-generated interpretation and regulation.

　　Bruce Schneier has noted that “when an AI takes over a human task, the task changes.” Actions taken by agencies to interpret statutes and regulations, and issue new regulations and guidance are no exception. Current AI models complete tasks in a different manner and commit errors differently than humans. These differences have legal significance. Respect accorded to agency decisions reflects the very human process of acquiring wisdom through the trial and error of administering the law.

　　Absent specific direction by Congress, the theories of judicial deference to agency actions generally do not mandate deference to actions exclusively or predominantly performed by AI. Use of AI models by agencies decreases the odds of those agencies being held accountable by the people, one of the central rationales for deferring to agencies over courts. AI models used by government officials are often privately owned and designed, which limits what the public can learn about those models. Agency officials may find AI as a useful scapegoat that misleads the public as to who bears responsibility for certain actions. Finally, AI works in a way that makes it difficult, if not impossible, to determine how and why it produces a certain result. A reality that exacerbates accountability concerns, may result in a head-first collision with the Administrative Procedure Act, and might undermine the rule of law.

　　AI models cannot apply “informed judgment,” in the sense of relying on years of expertise and experience to identify the proper regulatory response. This limitation arises because of two key factors. First, AI models are only as good as the data they are trained on and there is no guarantee that even models trained on specific information will be error free. Second, AI models process information and make decisions in a different way than humans.

　　Decisions made or significantly informed by AI also lack consistency. Unexplained and unpredictable changes characteristic of AI models conflict with the rule of law. The public may not receive adequate fair notice if regulations change at unexpected times and for unclear reasons.

　　Rather than wait for the moment of agency reliance on AI to interpret laws and regulations, scholars should clarify now how such AI interpretations ought to be treated by courts. The attention to administrative law following landmark Supreme Court decisions increases this sense of urgency. For a short while, admin law is headline news. That attention should not be squandered but should instead be channeled into further efforts to clarify the law and reinforce foundational principles of administrative and constitutional law.

　　This short essay does not answer all the questions raised by AI interpretations. For instance, what is the line between de minimis uses of AI to inform interpretations versus substantial reliance on AI to develop those interpretations? These and other questions merit further study. The difficulties associated with answering those questions, though, merit a default adoption of the Reality Doctrine. Unchecked encouragement of agencies using AI is misguided and irresponsible unless and until all legal actors understand how those uses fit into our broader systems of administration and judicial review.