The art of judging in the era of AI (Part 2)

A sign placed in 19th-century hotel rooms to introduce customers to electric lighting carried an instruction that now seems comic: "Do not attempt to light with match." The user was told instead to turn a key on the wall. The sign also assured customers that electric light was "in no way harmful to health" and would not disturb their sleep.

Those instructions made sense in their time. People were encountering a disruptive technology through the mental model of the then-familiar technology it was replacing. A lamp was something one lit with a match. And electricity presented genuine dangers as well as unfamiliar ones. Over time, we developed the knowledge, practices and standards necessary to use it safely, until something once novel became part of the infrastructure of everyday life.

Artificial intelligence is not electricity, and the analogy should not be pushed too far. But AI is another disruptive technology arriving faster than many of the practices needed to use it wisely. For the judiciary, that presents a problem of professional competence. It is not our first such challenge.

Many judges whose legal careers now span decades entered the profession during another technological disruption. Over time, computer-assisted legal research transformed how lawyers worked. Initially, Westlaw and Lexis were controversial. Law schools responded to that transition somewhat defensively: Even as we learned to conduct research electronically, we were still taught to "Shepardize" cases using physical volumes. I dutifully learned how. I have not done it since law school.

The lesson, however, was not wasted. The enduring professional skill was not how to turn pages in Shepard's. It was knowing that authority must be checked and exercising legal judgment about whether a case remains good law. Technology changed the method. It did not eliminate the responsibility.

AI presents that distinction on a much larger scale.

Today's law students are encountering AI while they learn to think as lawyers. Those of us who went to law school 20, 30, 40 or more years ago were trained to think as lawyers in a different technological environment. We were not taught how to frame a legal problem for an artificial reasoning system, how to interrogate its response, how to recognize when its framing may be influencing our own, or how to use it deliberately to challenge our analysis rather than simply confirm it.

Higher education is discovering that the problems posed by AI run deeper than teaching students how to operate a new tool. Researchers are asking how critical thinking itself should be taught when students can work with machines capable of participating in the analytical process. One recent framework for integrating large language models into higher education argues that unstructured use can encourage cognitive offloading and reduced "epistemic agency," while carefully designed instruction can instead preserve what the authors call "cognitive friction" and use of AI as a provisional thinking partner. (M. Vendrell and S.-K. Johnston, "Scaffolding critical thinking with generative AI: Design principles for integrating large language models in higher education," 10 Computers and Education: Artificial Intelligence 100572 (2026).) A systematic review published in the same journal synthesized 67 empirical studies and similarly concluded that the cognitive effects of ChatGPT depended substantially on how it was incorporated into the learning environment. (C. Li, H. Cui and L.S. Hagedorn, "The cognitive impact of ChatGPT in higher education: A systematic review of critical and creative thinking outcomes," 10 Computers and Education: Artificial Intelligence 100571 (2026).)

The implications for judicial education should at least make us curious. If universities are rethinking how critical reasoning should be taught in an AI-mediated environment, why should we assume that judges--many of whom learned their professional habits decades before generative AI existed--do not require corresponding adaptation?

In earlier columns, I wrote that courts had reached, and indeed passed, an inflection point with AI. That does not mean adjudicative AI is or should be ubiquitous. It means the technology is sufficiently capable and sufficiently present in legal practice that courts must make informed decisions about it.

One response is to exclude AI from adjudicative work. The risks of using AI provide substantial reasons for caution. AI can get the law or record wrong. More subtly, it can invite unconscious deference or make discretionary decisions appear more determinate than they really are, a phenomenon I called "discretion flattening" in an earlier article. (Karin Schwartz, "The art of judging in the era of AI," Daily Journal (Sept. 9, 2026).) But exclusion is a more difficult proposition than it first appears, and not merely because it is impractical. AI is already embedded in the practice of law that arrives at the courthouse, and it will only become more so. The meaningful choice is not whether AI will touch judicial work but whether it does so thoughtfully or haphazardly. We did not forbid electric light because it was unfamiliar and carried risks; we learned to use it safely.

Moreover, the case for engagement is affirmative, not merely a concession to necessity. Used well, AI can improve judicial work rather than simply accelerate it. It can stress-test a tentative conclusion, surface contrary authority, expose unstated assumptions, generate counterarguments and make a voluminous record more thoroughly reviewable. Yes, there is the risk of yielding too readily to artificial cognition, a phenomenon that researchers have described as "cognitive surrender."  (See 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).) The mirror image potential for AI, however, is cognitive amplification: using AI deliberately to extend and deepen, rather than displace, human reasoning. This is not wishful thinking, as the higher-education research noted above suggests.

That contingency should shape our response. If the danger lies in careless use and the value in disciplined use, the answer is to cultivate the discipline rather than ban the tool. We should not set the ceiling for a powerful technology at the level of its least careful user. Efficiency belongs in this account as well--not as mere convenience, but as a dimension of access to justice--since litigants are better served when courts resolve matters promptly and thoroughly. This does not mean AI belongs everywhere. As the model below reflects, sometimes the responsible answer is that a particular tool should not be used for a particular task. However, reaching an intelligent answer to that question, that is, identifying when AI is appropriate and when it is not, itself requires competence.

A more enduring response is to develop a model for ethical and responsible judicial use of AI, along with a pedagogy for teaching it. That model might begin with five stages:

AI literacy → tool selection → prompt design → evaluation and challenge → judicial decision.

Though presented as a sequence, these stages should be iterative rather than strictly linear. Evaluation and challenge may send the judge back to reframe the prompt or reconsider the choice of tool, and each pass deepens the literacy on which the first stage depends.

The first is foundational. Judicial AI literacy is not simply knowing that large language models hallucinate or learning a few prompting techniques. It requires understanding what different AI tools can and cannot do; what guardrails and safeguards are baked into a particular AI tool, and what additional ones may be user imposed; the risks associated with different uses; and how to use appropriate tools sophisticatedly to improve judicial work.

That last component matters. AI literacy should not be conceived solely as defensive knowledge. The same capabilities described above, such as the ability to interrogate an argument, expose assumptions and surface what a first pass missed, are as much a part of literacy as knowing where the tool fails. The same technology that can encourage passive reliance can, used differently, make the judge work harder.

Literacy permits the second step: tool selection. "Should judges use AI?" is too abstract to be particularly useful. The meaningful inquiry is whether a particular tool should be used for a particular task, and under what conditions. Sometimes the responsible answer is no.

If the answer is yes, prompt design becomes part of the intellectual work. What is the task? What information or authority should constrain the response? What assumptions should be tested? What should the system be asked to do, and not do? For judicial use of AI to be ethical and responsible, the initial development of the prompt may require a fair amount of work on the front-end. This is because a prompt is not merely an instruction to a machine; it frames an inquiry, and framing can affect the output and the additional human work that follows.

Next comes evaluation and challenge. Verification remains essential, but verification alone is not enough. A judge must test the analysis: What has been assumed? What has been omitted? Would another framing alter the analysis? Does apparent certainty conceal a range of permissible outcomes?

This is where AI may require us to develop new approaches to critical thinking. The danger is not confined to accepting an erroneous answer. An entirely plausible answer may anchor subsequent analysis, obscure alternatives or subtly narrow the field in which discretion is exercised. Conversely, a judge can deliberately use AI to expose those very problems, identify overlooked analytical possibilities, and generate counterarguments.

Finally comes the judicial decision. The technology may have helped research, organize, synthesize, test or challenge the analysis. But the judge must determine what remains a matter of judgment and exercise it.

None of these stages is revolutionary in isolation. Lawyers and judges have always had to understand their tools, choose appropriate ones, frame questions, evaluate answers and exercise professional judgment. What is new is the nature of the tool. AI does not merely retrieve information. It can participate in the analytical process itself. That makes sophisticated use potentially more valuable, and uncritical use more consequential.

It also means that judicial education cannot consist merely of presentations explaining generative AI, warnings about hallucinations and demonstrations of clever prompts. Some knowledge can be taught that way. Competence cannot.

Judges will need opportunities to use these systems, make mistakes, encounter persuasive but flawed analysis, recognize their own susceptibility to deference, learn to challenge outputs and discover ways AI can deepen rather than displace their reasoning. We may need to teach not only a new technology, but new critical-thinking practices for working with it.

The hotel sign told new users both what to retain and what to change: turn the key; put away the match. The challenge for the judiciary is considerably more complicated. We must identify which elements of our existing judicial craft are enduring, which practices reflected the technologies with which we learned them and what new competencies this disruptive technology requires.

The objective is not to produce judges who use AI. It is to produce judges capable of judging well in a world in which AI exists.

That will require more than rules. It will require education, experience and a pedagogy built for the task.

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.