An “I know it when I see it” approach to adjudication may work in some contexts but, for obvious reasons, will not in the case of existential risks: the stakes are too high to allow intuition to become the law of the land. Imagine, for instance, the following scenario:

     In the wake of ChatGPT6 being released and demonstrating capabilities beyond those even imagined by OpenAI CEO Sam Altman, U.S. Senators Marco Rubio (RFL) and Ron Wyden (D-OR) sponsor the “Responsible AI Development Act.”


     The Act defines “Frontier AI” as “any Artificial Intelligence model that may cause irreversible, significant, and widespread harm.” Furthermore, the Act authorizes the Secretary of the U.S. Department of Homeland Security to update that definition when necessary to prevent the occurrence of such harm.


     Microsoft and Google immediately challenged the constitutionality of the bill under the Nondelegation Doctrine, the Major Questions Doctrine, and the Due Process Clause. The Big Tech rivals turned co-plaintiffs allege that, as defined, “Frontier AI” is unconstitutionally ambiguous because it does not sufficiently identify the types of harms that may subject an AI model to additional regulation.


     A federal district court judge receives the opening briefs and quickly gets distracted by phrases such as “Floating-Point Operations,” “compute,” “foundation model,” and “fine-tuning.” The judge asks their clerks to explain these concepts. The clerks, also at a loss but eager to help their judge, turn to ChatGPT to help them understand these AI concepts. Unsurprisingly, the clerks return from their “research” more confused than informed. They update the judge on their findings (or lack thereof). The judge tells the clerks that they will all just have to “do their best” and “trust their guts” for the duration of the suit.

Problematic, right? Likely, yes.