"The Judicious Judge's Guide to Generative Artificial Intelligence and Large Language Models" advises members of the judiciary in areas such as basic technical and ethical guidance, responsible use of AI in chambers, AI issues in litigation and tasks for which the technology may be applied.
"If this article can help diminish the fear of the unknown, lower the blood pressure of folks who are uncertain about generative AI and have them just try it and make an informed decision, it will have accomplished a significant purpose," said co-author and recently retired Arizona Court of Appeals Judge Samuel Thumma.
The paper, which was released in late August and is set to be published in the Columbia Science and Technology Review in May, was written by Judge Thumma, U.S. District Judge Xavier Rodriguez in the Western District of Texas, U.S. Magistrate Judge Allison H. Goddard in the Southern District of California, Louisiana Fifth Circuit Court of Appeal Judge Scott U. Schlegel, and University of Waterloo and Osgoode Hall Law School professor Maura R. Grossman.
The guidelines in the paper are based on the American Bar Association's Model Code of Judicial Conduct.
Grossman said the guide is an expansion of a 2025 paper with guidelines on how to handle AI in the judiciary written by her and other members of an ABA committee. A year after it was written, she said, Judge Rodriguez reached out seeking to create a more comprehensive version of the paper, citing the need for constant updates to keep up with the fast-evolving nature of AI technology.
"We decided we were going to make it a much more practical guide that a judge could pull off their shelf when they needed to figure out things such as getting a prompt to do a certain thing or looking up guidance that states are giving," she said. "We wanted to make a comprehensive and self-contained guide that was very practical, because we were hearing from judges that there was nothing like that which they could go to."
The guide outlines a number of "ethically permissible" uses of AI in the courtroom, including database creation, legal research, drafting and editing, summarization of briefs, transcripts and evidence documents, court administration, translation and transcription.
Some of the key aspects of the principles governing generative AI use in the legal field include ensuring judicial authority remains human, establishing proper staff supervision and policies, and keeping decisions impartial and not using AI to make a final decision, according to the guide.
"We also encourage a 'sandwich approach' where you start by yourself, use AI in the middle as a tool to help you, and then you verify yourself," Grossman said. "We don't encourage starting a draft with an AI tool because it sort of locks you in to a position and can have certain kinds of biases that way. We encourage a judge to exercise their judgment first and only use these tools as supplements."
An appendix includes various sample prompts for how judges and their staff can put AI into practice. For example, prompts are offered for "editing and proofreading opinions," "hearing preparation" and "challenging your own initial assumptions." The latter reads in part: "You are reviewing the attached draft opinion as a skeptical colleague or appellate judge. Identify at least three aspects of the reasoning that might be challenged and explain the nature of each potential challenge. Do not suggest an alternative outcome; focus on analytic gaps."
Judge Thumma, now director of the Arizona Supreme Court's Center for Forensics and Artificial Intelligence after more than 20 years on the bench, said the appendix is useful for judges and courtroom staff to experiment with AI and increase their comfort levels with the technology, particularly given the range of opinions when it comes to judges and their own AI use.
Some judges don't use the technology, others are hesitant out of fear that it would make their jobs redundant, a large middle group is indifferent or intimidated by it, and another group has fully embraced it, Judge Thumma noted.
As the self-proclaimed "least tech-savvy person in his family," Judge Thumma said one of his goals is to have people become more comfortable with using AI technology through their experiences working in and navigating the legal system.
"There's room in these rooms for people who aren't computer programmers, who aren't tech-savvy, to really look at the technology and see how it can be accessible for those like me on the bench along with self-represented parties, lawyers and others who use the court system — improving the experience, not making things more complicated," Judge Thumma said.
The guide also contains multiple sections addressing AI-generated hallucinations, highlighting an issue that has become increasingly common in courtrooms across the nation, with attorneys frequently being admonished and in some cases sanctioned for legal documents they submitted.
"We obviously very strongly recommend that everything be verified, and some things are easier to verify than others," Grossman said. "Even with lawyers, it's not enough today to just check if the case exists; you have to actually check the quotes, and you have to check that you've reported it accurately."
Because the guide isn't expected to be published until May, Grossman said the authors already anticipate a number of updates and changes will have to be made by that point. Prior to its initial release, she said, they already had to include a section on prompt injection, which is when hidden instructions are included in a document that are invisible to a human but able to be understood by generative AI.
With the increased capabilities and prevalence of generative AI tools, the guide addresses their role in the increase in self-represented litigants and the potential pitfalls posed by improper AI usage in drafting documents such as discovery requests.
"It is likely that SRLs will use GenAI to draft discovery requests, and those GenAI-generated documents may seek nonrelevant, privileged or disproportional discovery," the guide said. "Another likely outcome will be the GenAI output may give the SRL an unrealistic assessment of the merits or monetary value of their claim. Judges should consider early intervention and status conferences to avoid discovery disputes."
Judge Thumma said that in his experience, the issue with hallucinations has been more prevalent with self-represented litigants than with lawyers in recent months, an observation he says makes him optimistic that attorneys are learning how to better use AI in their work.
"My hope is that it means lawyers are getting trained and educated, checking the citations they provide the courts," he said. "A self-represented party doing that is understandable, because if you have nothing and are just looking for some help, generative AI pulling something together that apparently reads in English adds value to a self-represented litigant."
Among the overarching goals behind the guide is to help make the legal system more efficient. Given how common it is for court proceedings to have a slow pace and long delays, especially in state courts, many people never end up getting justice, and one of the key goals of encouraging AI adaptation is to expedite the legal process, Grossman said.
"Anything that helps speed up the process and can increase efficiency, especially with repetitive tasks where judges don't need to exercise a lot of judicial thinking, is helpful," she said. "For example, if something comes in that's handwritten, transforming it into a typed version makes things much easier to deal with. AI is a remarkable help with these types of tasks."
Grossman said that going forward, while she can envision a future in which certain types of legal matters such as small claims are handled completely by AI technology, she expects the more human aspects of law, such as criminal and family matters, will continue to be handled by human judges. The ancillary parts of those legal areas are where she said she does expect there to be a major push for AI usage.
Judge Thumma added that he fully expects AI to become as ubiquitous as the personal computer, with the main question being the form in which it's ultimately used.
At the same time, he acknowledged, many concerns still exist, and the technology isn't a complete solution for everything.
"It can do incredible things and I do think it will become pervasive in a variety of different ways while bringing a whole bunch of other issues along with it," Judge Thumma said. "I am not a person who views generative AI as a panacea, and there are downsides. We need to figure out guardrails, and that's hard, because the technology is advancing so quickly along with the environmental cost. I don't view it as an answer to everything without pause or concern, but I also do think its use will continue to expand."
--Editing by Alanna Weissman.
The forthcoming paper is available through Social Science Research Network, or SSRN, which is owned by Elsevier and is part of the RELX Group, which also owns Law360.
For a reprint of this article, please contact reprints@law360.com.