The art of judging in the era of AI

Several weeks ago, I posited that our conversation about legal AI appropriately began with hallucinations but should not end there. (Karin Schwartz, "Beyond hallucinations: Is it time to recalibrate?" Daily Journal (July 22, 2026).)  As legal AI improves, fabricated cases and quotations may become less important than subtler errors of legal analysis: a missed distinction, a mischaracterized holding, an overlooked nuance.

There is another, more subtle, and more profound risk to talk about. Unlike hallucinations or faulty legal analysis, it does not require the AI to be wrong. I will call it discretion flattening.

Judges routinely decide questions for which the law does not dictate a single answer. Sometimes statutes or rules expressly confer discretion. Sometimes courts must balance multiple factors. Sometimes the governing standard--reasonableness, proportionality, good cause, prejudice, the interests of justice--necessarily leaves room for judgment. Within legal boundaries, more than one outcome may be defensible. Sometimes judicial discretion may lie in which remedy(ies) to order, and how narrow or broad to craft any final order or judgment.

Discretion is where much of the art of judging occurs. Now introduce increasingly sophisticated legal AI. And add in the human inclination to sometimes defer.

An AI system may accurately identify the governing authorities. It may correctly summarize the relevant facts. It may organize competing arguments, identify the applicable factors and synthesize all of that information into an impressively coherent analysis. It may then recommend an outcome, or even set of outcomes, that is (or are) entirely reasonable and legally supportable.

Nothing has hallucinated. Nothing is obviously wrong. Yet, something important may nevertheless have happened.

The synthesis may make the recommended result appear more inevitable than it is. Competing considerations may have been compressed into a clean analytical path. A range of permissible outcomes may subtly become a preferred outcome. Judgment may begin to look like calculation.

That is what I mean by discretion flattening: the potential for AI-assisted analysis to obscure, narrow or compress the space in which judicial discretion exists and should be independently exercised.

The phenomenon can operate at more than one level. At its most basic, AI may flatten the perception of discretion. A judge presented with a polished analysis may fail to recognize that the law leaves more room for judgment than the analysis suggests.

More subtly, AI may flatten the exercise of discretion. The judge may fully understand that discretion exists but become anchored to the AI's framing of the problem or its proposed resolution. The judge still decides. The judge may even reach a legally supported and analytically justifiable result. But the range of possibilities may have been cognitively narrowed before the judge begins independently engaging with it.

Consider a generic judicial problem. The law identifies several factors and gives the court discretion to balance them. Some point one way, some another. Reasonable judges could weigh them differently.

A capable legal AI may synthesize those factors and explain why, on balance, one result is preferable. That may be both useful and efficient. But the very coherence that makes the analysis useful may also conceal something important: there was no algorithmically correct way to assign relative weight to those factors in the first place. The weighing was the judging.

A skeptic might object that there is nothing new here: A forceful brief or a persuasive bench memo can compress a judge's sense of the available options just as easily. The difference is the source. What is distinctive about AI is that the compression now arrives clothed in the authority we tend to accord machines, and, as research suggests, such analysis is often trusted more, not less, than the same reasoning offered by a person.

This concern does not arise in a cognitive vacuum. In 2019, for example, in a study that predated the rise of AI as we now know it, researchers demonstrated through a series of experiments that humans adhere more to advice when they think it comes from an algorithm than a person. (Jennifer M. Logg, Julia A. Minson and Don A. Moore, "Algorithm appreciation: People prefer algorithmic to human judgment," Organizational Behavior and Human Decision Processes (March 2019).)  

More recently, researchers at the Wharton School coined the term "cognitive surrender" in relation to the use of AI. (Steven D. Shaw and Gideon Nave, "Thinking--Fast, Slow and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender" (The Wharton School), available at https://osf.io/preprints/psyarxiv/yk25n_v1 (working draft, accessed Sept. 2, 2026).) Building on the familiar distinction between intuitive and deliberative human reasoning, the authors posit a third system: artificial cognition operating outside the human mind. Their research identifies cognitive surrender as the adoption of AI outputs with minimal scrutiny, displacing the user's own intuitive and deliberative reasoning.

Across three preregistered experiments, the Wharton researchers found that access to AI substantially improved performance when the AI was right--and reduced performance when it was wrong. Perhaps more troubling, interacting with AI increased participants' confidence, even following erroneous AI advice.

Discretion flattening is not the same phenomenon. Nor does cognitive surrender necessarily cause it. But the possibility of cognitive surrender may amplify its consequences.

A judge confronted with obviously defective AI output has a reason to become skeptical. A judge confronted with excellent AI output may have less reason to do so. And as legal AI becomes more capable, its greatest influence on judging may therefore come not when it fails spectacularly, but when it succeeds persuasively.

That matters because judging is not simply the production of legally permissible answers. We entrust decisions to judges, in part, because many legal questions cannot be resolved mechanically. Judges are expected to identify the boundaries established by law and then exercise judgment within them. That requires context, proportionality, experience, an appreciation for competing interests and, sometimes, recognition that two judges applying the same law conscientiously could reach different conclusions.

There is an irony here. Much of the current debate about judicial AI appropriately emphasizes that AI must remain subordinate to human judgment. India's recently proposed Regulations for Use of Artificial Intelligence in Courts, for example, expressly place "human primacy and judicial independence" at the center of their framework and provide that AI should remain assistive rather than supplant independent judicial authority. (See https://cdnbbsr.s3waas.gov.in/s3ec0490f1f4972d133619a60c30f3559e/uploads/2026/06/2026060342.pdf, accessed Sept. 2, 2026.)

But preserving human authority formally is not necessarily the same thing as preserving human judgment functionally. A judge can remain the final decision-maker and still be influenced by the way an AI has defined the decisional landscape. A human signature at the bottom of an order does not tell us whether the human recognized the full scope of the judgment the law entrusted to her.

This is why discretion flattening deserves attention even if legal AI becomes extraordinarily accurate. Perhaps especially then.

The danger is not simply that AI could cause judges to reach bad decisions. It is that, without recognizing what is happening, judges could gradually become less attentive to the places where judging rather than legal analysis is required. Over time, that could affect not merely individual outcomes but the quality of judicial decision-making itself.

And ultimately, legitimacy is implicated. Public confidence in courts depends on more than technical legal accuracy. Our system entrusts judges with authority because we expect them to exercise independent judgment. If increasingly capable machines participate in judicial analysis, preserving that judgment requires more than ensuring that a human remains nominally "in the loop."

It requires preserving the art of judging.

In an earlier article, I wrote that we appeared to be approaching an inflection point in the judicial use of AI. (Karin Schwartz, "Beyond chatbots: Is there a role for adjudicative AI?" Daily Journal (July 28, 2026).)  I am now persuaded that we have passed it. That does not mean most judges are extensively using AI for adjudicative purposes. They are not. Nor does it mean that every possible use of adjudicative AI is desirable. Some may not be.

It means something narrower: Courts can no longer treat the effect of AI on judging as a hypothetical problem for another day.

Recent developments make that increasingly apparent. India is already constructing a comprehensive framework for AI in its courts. Recent Japanese research has examined public acceptance of AI adjudication across different kinds of disputes, finding meaningful differences between institutional or procedural disputes and those involving interpersonal relationships and emotional complexity. (Masahiro Fujita and Eiichiro Watamura, "Psychological Features of Dispute Content and Public Acceptance of AI in Legal Adjudication: Evidence for Systematic Variation Beyond Individual Differences" (March 10, 2026), working draft, available at https://ssrn.com/abstract=7017378.)

The questions are no longer confined to whether AI might enter adjudication. Institutions are beginning to confront where it belongs, where it does not and what meaningful human judgment requires when it does.

Discretion flattening is one piece of that larger inquiry. Naming the phenomenon is only a beginning. If judges are to use increasingly capable AI without diminishing the very capacity for judgment that justifies their role, we will need more than rules telling judges to retain responsibility. We will need to understand what responsible judicial engagement with AI actually looks like, and how judges can be trained to do it.

That will be the subject of the next article.

Author's Note: In the interest of transparency, I used legal AI tools during the research and development of this article. I independently reviewed the authorities. I accept full responsibility for every statement and conclusion expressed here.