Insights from The Judicial Perspective on AI and Legal Technology at Everlaw Summit ‘26
Three members of the bench weigh in on where disclosure rules fall short, how chambers are actually using generative AI, and the responsibilities no judge should ever delegate.
by David Pemberton
Speaking to a room of litigators this fall during the Everlaw Summit ‘26 judges panel, three members of the bench shared their candid perspective on how artificial intelligence and emerging legal technologies are reshaping the judiciary.
Moderated by Everlaw Chief Legal Officer Gloria Lee, the panelists Hon. Allison H. Goddard (S.D. Cal.), Hon. Willie J. Epps, Jr. (W.D. Mo.), and Hon. Samuel A. Thumma, recently retired after twenty years on the Arizona bench and now director of the Arizona Supreme Court's Center for Forensics and Artificial Intelligence, focused their discussion on how courts can balance security with flexibility by building practical guidelines inside chambers, extending mutual grace when errors happen, and preserving human accountability as new technologies enter judicial workflows.
Judges Meet Rising AI Adoption with Curiosity and Scrutiny
The panel began with a look back at how generative and agentic AI capabilities have affected legal work in the past year. When asked how artificial intelligence has reshaped the judicial system, Judge Goddard said: “From my perspective, what I see is a really welcome change, that a lot more judges are interested in learning how to use AI.”
While growing judicial enthusiasm points to AI's potential, the technology's rapid adoption by lawyers has also introduced new verification challenges for the bench.
Judge Epps noted that over the past year, his approach to reviewing court filings has shifted toward much heavier scrutiny because of AI’s proliferation in legal drafting. Rather than taking filings at face value, Judge Epps said he and his staff now examine submissions "with a critical eye, even from the top law firms." He pointed out that major firms frequently employ junior attorneys who are still learning the ropes and may not be "supervised as well as they should be," making it essential to verify that cited cases actually exist and genuinely support the legal arguments being made.
That same heightened vigilance extends inside his own chambers, where incoming law clerks are likewise still finding their footing. Judge Epps emphasized that he reviews their drafts far more critically than in the past now that they are using AI, so that he can ensure every citation is accurate. Ultimately, his motivation comes down to protecting the integrity of his bench. "I just don't want to be a judge you read about who let something out of chambers that is just dead wrong," Judge Epps said, adding, "That would be incredibly embarrassing. You all have seen those headlines and I don't want to be that guy."
Echoing that need for clearer boundaries and training, Judge Thumma compared the arrival of generative AI in chambers to the early days of social media, recalling an intern who casually mentioned sharing her courtroom experience with her online followers, a moment that underscored the importance of setting expectations for new technologies on day one.
While there are still judges who "wish that generative AI didn't exist because in their life it doesn't," Judge Thumma explained "a tide is rising" of colleagues who are curious about AI and have begun using the technology in increasingly sophisticated ways. To move beyond anecdotal impressions and establish a baseline for future training, Arizona recently distributed a statewide survey across its appellate and trial benches to gauge how judges are actually using AI and what guidance they still need, an effort Judge Thumma hopes other courts will replicate "to see what needs to be done going forward."
While internal training and surveys help judges get their own houses in order, the bench has also had to decide how to oversee the rapidly expanding use of AI by litigators.
Weighing the Value of AI Standing Orders in the Courts
One of the most prominent responses to that challenge has been the rise of judicial standing orders requiring lawyers to disclose or certify their use of generative AI. During the panel, the judges weighed whether these rules actually help prevent errors, or simply create unnecessary friction.
Lee pointed out that lawyers are currently navigating a fragmented patchwork of "884 different standing orders or local rules for disclosure of generative AI use" across federal and state courts nationwide.
"These standing orders are not necessary," Judge Goddard said. "I've been telling judges all around the country, they're not necessary." To illustrate how easily broad restrictions can backfire, Judge Goddard looked to a recent standing order banning attorneys from using artificial intelligence to draft any court brief. Because the order doesn’t use the term “artificial intelligence” without specifying "generative AI,” if taken literally, Judge Goddard quipped, such a rule would leave lawyers "handwriting their stuff and then not scanning it in," since everyday legal software relies on AI.
"And do you think that makes us look good? No. It makes us look like we don't understand the technology," Judge Goddard continued. “We have ways to address the hallucinations, and the hallucinations, they’re not great, but they’re easy to detect, they’re easy to address.” Judge Goddard suggested that courts should "refocus on the tools we already have to address these issues."
Beyond their practical limits, Judge Thumma questioned whether disclosure rules make sense for judicial review at all. He asked whether a judge would realistically apply extra scrutiny to catch fake citations only when a filer admits to using generative AI, noting that courts should apply "that same degree of scrutiny for everything that comes their way."
Judge Thumma concluded that standalone AI orders fail to move the needle and risk undermining public confidence in the bench. "So I don't think they accomplish anything," he said. "I think at times they're almost an embarrassment and I don't like the message that they send to the world about what we're doing."
Practical Chambers Guardrails Outperform Blanket Bans
Of course, moving away from formal standing orders doesn't mean taking a hands-off approach. Instead, the panelists emphasized building practical guardrails and shared fluency inside their own institutions.
Judge Epps brings that pragmatic openness to the judicial writing course he teaches at the University of Missouri, telling students on day one, "I welcome your use of AI." Judge Epps warned that treating AI like a temporary fad, much like early skeptics dismissed email and the internet, will leave legal professionals behind. "I don't think you can be a strong judge if you don't understand this technology, if you're not working with it, if you're not trying to get better and stronger at it, I think the world will pass you by."
I don't think you can be a strong judge if you don't understand this technology, if you're not working with it, if you're not trying to get better and stronger at it, I think the world will pass you by.
To help more cautious colleagues move past hesitation and see how AI fits into daily judicial workflows, Judge Goddard’s district formed a working group she calls "the trifecta," bringing together judges, law clerks, and IT. Because producing court orders relies heavily on administrative staff and clerks rather than judges alone, the group—which has since become a model for other courts—brainstorms ways to use AI "not just to be faster, but to be better." Jokingly admitting that "I write in crayon and they write in pen when it comes to AI," Judge Goddard praised her clerks' sophisticated prompts for handling judge-tailored proofreading and drafting scheduling orders.
Figuring out how to support safe innovation for judges without legal staff at their disposal is especially urgent given the risks of uncoordinated experimentation. Judge Thumma warned against the danger of a judge "going rogue" with off-the-shelf tools simply because they seem convenient, without weighing confidentiality, personally identifiable information, or cybersecurity threats.
Beyond data security, Judge Thumma stressed the need to keep a "human in the loop" so technology never displaces judicial reasoning. "Is the judge making the decision or is the bot making the decision?" Judge Thumma asked. "In our world right now, we're not ready for the latter."
Unchecked AI Hallucinations Carry Severe Professional Consequences
One week before the panel, the Fifth Circuit supplied a stark example in Jackson Federation of Teachers v. Fitch, a high-profile lawsuit challenging Mississippi House Bill 1193. A district judge had initially issued both a temporary restraining order and a preliminary injunction blocking the law, only for both to unravel over AI-generated fabrications including non-existent plaintiffs, unrecorded witness declarations, altered statutory text, and a citation to a completely fabricated precedent.
Following an inquiry, the judge acknowledged that a law clerk had used generative AI to draft the order, though he initially tried to downplay the mistakes as minor clerical errors before issuing a replacement order that still contained a hallucinated case. A unanimous Fifth Circuit panel agreed that the lower court's handling of the matter was a "botch job,” and took the rare step of removing the judge from the case entirely on remand.
Judge Thumma had read the Fifth Circuit's opinion the night before and predicted the decision "will become infamous." Beyond reversing and remanding the ruling, the Fifth Circuit panel barred the judge from ever touching the case again, ordering the chief judge to reassign it. Reflecting on his own years on the trial bench, Judge Thumma called that outcome "a very bad day," adding, "I mean, that's about as bad as it gets professionally."
"As a federal trial judge, reading what the Fifth Circuit wrote just sent chills through my body because every federal trial judge in America has been there," Judge Epps added, reacting to the appellate rebuke. He noted that the situation worsened because "the trial judge replaced the document and the replacement had hallucinations as well, and the Fifth Circuit ripped him a new one for that."
While emphasizing he was "not making excuses for that judge," Judge Epps explained the human realities behind chambers workflows: judges operate in a collaborative environment where they rely heavily on law clerks, and no one outside chambers knows "what kind of day the judge was having" when he reviewed the initial draft or signed off on the flawed replacement. In other words, these things happen.
Extending Grace in a Collaborative Legal System
Judge Goddard noted that many recent sanctions orders for hallucinated citations fail to reflect "the workflows of the modern law firm," and that some courts have overreacted by penalizing every attorney listed on a pleading's caption even if they didn't sign the filing, a trend already prompting lawyers to keep their names off captions altogether.
Drawing a parallel to how chambers function, Judge Goddard explained that judges rely on a similar "honor system" with their own law clerks, since it would be impossible to issue rulings on time if judges had to manually verify every single citation clerks provide. Because courts understand those collaborative pressures, Judge Goddard suggested they should extend "a little bit of grace,” though she acknowledged that in the Fifth Circuit case, "there was no more grace left to be given."
Building on that theme, Judge Epps said that AI-driven errors—ranging from naming "the wrong parties that weren't part of the case" to inventing "made up quotes from the record that just didn't exist"—are already happening quietly in local courts without making national headlines.
In recent instances involving solo practitioners, several of his colleagues held quiet hearings and have "shown grace to the parties and tried to protect the lawyer who has made those mistakes," recognizing that big law firms have fallen into the exact same traps. Addressing the practitioners in the room, Judge Epps asked attorneys to resist the urge to pounce on an opponent's misstep, reminding them that "you may have an opportunity to show grace at some point" and warning against the impulse "to stick it to the other side" when anyone could make a similar mistake down the road.
Human Judgment and Final Accountability Remain Non-Negotiable
To close the panel, Lee posed what she called "the million-dollar question everyone's struggling with": as AI grows more capable, how should courts decide what tasks to delegate to technology and what must remain strictly human? Lee challenged the panelists to pinpoint how the bench can preserve "that essential human judgment that is at the heart of adjudication."
Judge Thumma answered first, explaining that while courts have issued plenty of high-level guidance on approved platforms and general use cases, practical, ground-level standards are still lacking. "This is complicated stuff and there's less guidance on this than there should be," he said, adding that he hopes to help fill that gap through Arizona's Center for Forensics and Artificial Intelligence. Assuming confidentiality and ethical rules are satisfied, Judge Thumma argued that the clearest way to draw the boundary is by defining what generative AI should never do.
At the top of his list is anything involving credibility determinations, where trial judges watch witnesses "sweat and blink and pause" in real time. Weighing conflicting evidence and exercising judicial discretion shaped by years of lived experience on the bench are equally off-limits, areas he called "a non-starter" for AI. Because appellate courts review legal rulings rather than making initial factual findings, Judge Thumma noted that appellate work naturally lends itself more to AI assistance than trial work, particularly on purely legal questions like a motion to dismiss for failure to state a claim. Yet regardless of how much a tool assists with the legal analysis, he insisted that final accountability never shifts: "The judicial officer selected, elected, appointed, confirmed—their name needs to go on the bottom of the page, they can't race away saying, 'Oh no, no. Well, that was a generative AI tool, not me.'"
We need to remember that adjudication isn't just the job of one person. It's a whole system that works together.
Building on that human-centered focus, Judge Goddard broadened the perspective, emphasizing that judicial decision-making relies on a collective ecosystem rather than an isolated piece of software. "I would just say we need to remember that adjudication isn't just the job of one person," she said, pointing out that attorneys, witnesses, and the full evidentiary record all feed into the process. "It's a whole system that works together." For Judge Goddard, keeping the focus on the courtroom itself serves as the ultimate safeguard against over-reliance on automation, reminding the bench that judging is far more complex than "putting something in a tool and getting a result spit out."
Judge Epps largely agreed, viewing the debate through the lens of how higher courts will eventually weigh in on these issues. Still, he pushed back on the idea that AI's role will remain permanently limited. Even if the technology isn't quite there today, Judge Epps warned that Judge Thumma's current outlook could look "way, way off base" within five years as AI capabilities rapidly evolve. Though, in the end, Judge Epps concluded that courts "should not yield to AI."
Keeping Justice Human in the Age of AI
The Everlaw Summit ’26 judges panel made clear that the future of AI in the courts won’t be decided by a patchwork of local standing orders or blanket bans. As tools grow more capable and caseload pressures continue to mount, judges and practitioners alike will inevitably lean on AI to work more efficiently. Yet no matter how sophisticated the software becomes, the core mechanics of justice remain distinctly human: rigorous verification, mutual grace across the aisle, and an unshakable line of personal accountability. Technology can assist the work of the court, but it will never sign its name at the bottom of the page.
David Pemberton is an associate content marketer at Everlaw. His writing explores the influence of emerging technologies on the practice of law. See more articles from this author.