Shlomo Klapper (United States Court of Appeals for the Second Circuit; Yale University, Law School, Students) has posted The Responsive Court: AI, Legitimacy, And The Future Of Adjudication on SSRN. Here is the abstract:
Two forces are converging on the American judiciary: collapsing public confidence and the operational arrival of the AI judge. The response has calcified into an impasse. Functionalists treat judging as information processing and conclude that AI does it better. Purists insist something essential would be lost but offer no path through the crisis they refuse to solve.
The debate is stuck because both sides share an unexamined premise: that judging is one thing. It is two. Judicial work confronts epistemic uncertainty, which admits of resolution through information (what the contract says, when the accident occurred), and constitutive uncertainty, where the answer does not exist until someone with authority creates it (whether this sentence is just, whether this parent is unfit). Constitutive judgment requires moral agency: the capacity to stake one’s identity on a contestable value choice and take responsibility for the violence that follows. Each camp is right about different things and wrong to generalize.
This Article proposes responsive adjudication: a mechanism design for courts that calibrates human involvement to the nature of the judicial task. Drawing on Ayres and Braithwaite’s responsive regulation and the fiftieth anniversary of Sander’s multi-door courthouse, it develops a trigger architecture matching process to each case’s epistemic or constitutive character, introduces responsive legitimacy as a distinct category of institutional legitimacy, and demonstrates that the Supreme Court’s due process jurisprudence from Mathews v. Eldridge through Hamdi v. Rumsfeld has been groping toward this framework for fifty years without the conceptual vocabulary to reach it. Wherever decisions are made under authority, the same question applies: what requires finding, and what requires judgment? Let machines handle the first. Let humans answer for the second.
Recommended.