This is the second in a series on AI and the future of legal judgment. The first article, “The AI End of Integrity.” examined what is lost when accountability is removed from adjudication. This piece examines a separate but related question: whether an AI adjudicator is in fact as neutral and impartial as it is assumed to be.
The argument for AI adjudication rests on a foundational assumption: that an algorithm, unlike a human judge, is free of bias. Remove the person from the bench, the reasoning goes, and you remove the problem. No fatigue, no personal history, no unconscious preferences. Just the law, applied cleanly to the facts.
It is a compelling assumption. But it is wrong.
An AI adjudicator would be biased. Bias is built in. Built in to the algorithm by the people who design it and then by the training data. (“The Emperor’s New Codes: Seeing Bias in the Algorithmic Age” on TechCred). The important question is whether its bias would be better or worse than a human judge’s and whether the systems we rely on to catch and correct judicial error would have any purchase on it at all.
Where AI Bias Comes From
AI systems acquire bias through several distinct paths, and none of them are unique to consumer applications or search engines. They would apply with equal force to any AI system trained to make legal determinations.
The first pathway is training data. An AI learns from what it is fed. In the context of legal adjudication, that means decades of judicial decisions. That includes disparities that arise not from injustice but from the natural demographic and geographic character of the jurisdictions where those decisions were made. An AI trained on American case law would encode, at a structural level, any disparity present in those outcomes. It would not introduce new bias so much as formalize and perpetuate past bias, presenting historical inequity back to us dressed in the language of algorithmic neutrality.
The second pathway is design. Every AI system reflects choices made by its designers: what data to include, what outcomes to optimize for, how to weight competing factors. Those choices are not neutral. They are made by people with perspectives, priorities, and blind spots. They are also shaped by the cultural assumptions of the moment and place in which they are working — assumptions so embedded they are rarely visible as assumptions at all. The fact that those choices are embedded in code rather than expressed in a judicial opinion does not make them less consequential. It just makes them even less visible.
The third pathway is feedback loops. AI systems that are evaluated based on how closely their outputs match existing legal outcomes will, over time, become very good at replicating those outcomes, including the ones that reflect the patterns we might most want to correct. Optimization, in this context, is not the same as improvement.
The Scale Problem
A human judge’s bias is individual. It affects the cases that come before that judge at a particular point in time. It can be documented, appealed, and in extreme cases, raised as grounds for recusal. It is, in the most literal sense, localized.
An AI adjudicator’s bias is systemic. A single model, deployed across a jurisdiction or a court system, applies identical assumptions to every case that passes through it simultaneously. An error in the model is not an isolated incident. It is the same error, reproduced perfectly and at scale, in every courtroom at once.
Lawyers who practice across multiple jurisdictions understand instinctively what this means. Diversity in judicial interpretation, the fact that different judges bring different perspectives to the same legal question, is not a flaw in the system. It is its strength and the force behind how law develops. Outlier decisions create friction. Friction generates appellate review. Appellate review builds doctrine. Remove the variation and you remove one of the primary engines of legal evolution.
What Balance Means
When I think about balance I think about a tightrope walker, I see the perfect embodiment of the performance of balance. But they are not balanced. They are continuously, actively, effortfully correcting. The appearance of effortless stillness and neutrality is the product of experience, highly skilled, constant micro-adjustment, of a body that is always on the verge of failing and compensating in real time.
That is exactly what a human judge is doing. And it is exactly what an algorithm cannot do. Because an algorithm doesn’t truly correct, it optimizes. It doesn’t feel the wobble, and it doesn’t know it’s falling because it interprets the tilt as ground truth, and then reproduces the error at scale across ten thousand cases without that second thought.
The tightrope walker also embodies the mortality dimension—consequence, weight, what is at stake for the person in the room. Life. Someone on a tightrope knows what falling means. That knowledge is in their body. It is inseparable from how they move and why they move.
The Accountability Gap
Due process is a right to be heard by a decision-maker whose reasoning can be examined. An opaque algorithm cannot satisfy that requirement. "We cannot fully explain how the model reached this conclusion" is no answer, even if it is the only honest one available.
Human judgment operates inside a system designed to catch its failures — through a diversity of decision-makers, and through recusal, appellate review, and judicial conduct boards. Because no single perspective dominates, the system remains resilient. These mechanisms exist because bias in human judges is real but is recoverable. AI adjudication changes that calculus entirely. Training data encodes historical patterns, which litigants prevailed, which arguments succeeded, which outcomes followed which facts, carrying within them everything the legal system has ever gotten wrong. A single model deployed at scale replaces that diversity with singularity. What would be an occasional failure in a single courtroom becomes structural across every proceeding the system touches. The accountability mechanisms were built for an exception. At scale, the exception becomes the rule and there is no mechanism for holding the rule to account.
And there is a deeper problem. A human judge is part of the world the law is trying to regulate. They feel shifts in context emerging in real time. They can distinguish the subtle difference between a legal landscape that looks the same on paper but has shifted underneath. That permeability is not a weakness. It is how law evolves over time. An AI system cannot feel the ground moving. It applies yesterday’s categories to today’s conditions with perfect consistency and no awareness that consistency without context has become the problem.
What Should We Be Asking
AI used appropriately in research, in document review, in pattern analysis that supports human decision-making has enormous potential.
The question we should be concerned with is not whether an AI can apply the law more efficiently than a human. It is whether efficiency is the right aim for a system whose deeper purpose is justice.
An efficient system that encodes bias at scale and shields it from challenge is no improvement at all.
We must keep humans in the room – not because they are perfect, but because our awareness of our own fallibility is the very thing that forces us to hold the line.
I am grateful for the deliberately provocative clip that inspired this Continuance. It is no longer available.