AI Adjudication Debate Turns on Normative Uncertainty, Not Just Human Oversight

Courtney Cox’s forthcoming paper challenges policymakers to define what counts as a better legal decision before treating explainability or human review as sufficient safeguards.

A new discussion of Courtney Cox’s forthcoming paper, Hardwiring Hercules?, argues that policy debates over AI in legal adjudication may be asking the wrong first question. Rather than focusing only on whether people have a right to a human decision, Cox contends that regulators and courts must confront normative uncertainty: disagreement or uncertainty about what the decisionmaker ought to do in the first place.

Cox, an associate professor at Fordham Law School, discussed the paper with Kevin Frazier, an AI Innovation and Law Fellow at the University of Texas School of Law and a Lawfare senior editor, in a Lawfare episode published July 21.

The argument reaches beyond the familiar concern that an automated system may be unable to explain how it reached a result. AI systems can have difficulty supplying meaningful reasons for an adjudicative outcome, but human judges can face a related limitation when the governing legal or moral answer is uncertain. A human decisionmaker may offer reasons without resolving whether those reasons identify the correct outcome.

That distinction matters for compliance teams building or procuring AI-assisted systems for legal, administrative, benefits, employment, licensing, or other high-stakes decisions. A requirement for human review may establish accountability and an escalation path, but it does not by itself establish a decision standard. If the reviewer lacks a clear account of what result the law requires, substituting a human for an automated system may move the uncertainty rather than resolve it.

Cox’s paper also challenges a common alternative to a human-decision right: the claim that affected people are entitled to a better decision, regardless of whether a person or machine makes it. That standard appears technology-neutral, but it depends on a prior and difficult question: better according to whom, and measured against which legal, institutional, or ethical criteria?

For AI governance, the practical consequence is that decision-quality rules cannot stop at accuracy metrics, error rates, or explanation requirements. Those controls can test whether a system performs consistently against a defined benchmark. They cannot independently determine whether the benchmark captures the right interpretation of a contested rule, the appropriate balance between competing values, or the proper exercise of discretion.

This creates an enforcement challenge. Regulators can require documentation, audit trails, notice, appeal rights, human review, and explanations. Those obligations make automated adjudication easier to inspect and challenge. But they do not settle the normative choices embedded in eligibility rules, risk scores, evidentiary thresholds, or prioritization systems.

Organizations deploying AI in adjudicative workflows should therefore separate two compliance questions. The first is procedural: can the organization show how a system reached a recommendation or outcome, identify responsible personnel, and provide a meaningful route to contest the decision? The second is substantive: has the organization specified what a legally and normatively sound outcome looks like where the applicable standard is ambiguous?

The second question is harder, but Cox’s framing suggests it cannot be avoided through a simple human-in-the-loop requirement. Human oversight can be important when automated tools affect rights or benefits, yet oversight must include a defined authority to disagree with the system and a documented basis for resolving uncertain cases.

The Lawfare discussion confirms that Hardwiring Hercules? is forthcoming, but does not establish whether the full paper has been formally published or provide publication details. It also does not set out a specific regulatory proposal. Its immediate significance is conceptual: AI adjudication governance may need to address not only who makes a decision and whether they can explain it, but also whether institutions have articulated what the decision should be.