Two of California’s largest trial courts are testing an artificial intelligence tool that can help draft judicial orders and produce research memos, raising questions about transparency, accuracy and whether litigants should know when AI is being used in cases that affect them.
So far, judges have used the technology mostly in civil matters. But records obtained by CalMatters show the tool could eventually be tested in criminal proceedings, where decisions can affect a person’s freedom and access to justice.
The Los Angeles County Superior Court began a pilot program in February using software developed by Learned Hand, a company founded by Shlomo Klapper. The Riverside County Superior Court also has a $10,000 agreement with the company to test the program, mostly for civil and probate research memos that assist judges in decision-making.
Learned Hand uses large language models from Anthropic, OpenAI and Google to function as a virtual AI assistant for judges. The company says it tests the technology for bias and accuracy, though it has not publicly released those results.
In Los Angeles County, the Superior Court’s contract is worth about $314,000 and includes a roadmap for possible testing in criminal, family and probate divisions. Court officials did not provide CalMatters with detailed criteria for deciding whether the tool could safely be expanded into criminal and family courts, where the stakes are often far higher than in routine civil disputes.
One Los Angeles County Superior Court judge, who spoke to CalMatters on condition of anonymity because of judicial conduct rules, said the possibility of using the technology in sensitive criminal matters was alarming. The judge said colleagues discussed during a recent lunch that the tool could one day be used to help review appeals from people who argue that racial bias affected their conviction or sentence. California courts have seen a surge in such claims since lawmakers passed the Racial Justice Act in 2020.
“I find it outrageous,” the judge said. “AI cannot and will never replace human judgment when it comes to evaluating complex social dynamics. Ultimately, this will erode public confidence in the competence and fairness of the judiciary.”
Documents obtained through public records requests show most California superior courts now have policies governing generative AI. The state Judicial Council required courts to adopt such policies before using the technology. Of the 51 courts that responded to CalMatters, about a dozen said they use AI tools from companies including LexisNexis, Thomson Reuters and Microsoft Copilot.
The use of AI in courtrooms remains controversial because large language models can generate false citations, misleading summaries and overly confident conclusions. A 2025 MIT study found that models from major companies, including Google and Anthropic, can reduce critical thinking and brain activity.
AI-generated errors already have appeared in legal proceedings. Researcher Damien Charlotin has documented hundreds of instances in which litigants, attorneys and judges made mistakes after relying on AI, including nearly 90 cases in California-based state or federal courts since August 2024.
Last fall, a Los Angeles attorney was fined $10,000 for citing nonexistent cases. Earlier this month, The Sacramento Bee reported that AI use led to errors in four cases handled by prosecutors in Nevada County. Most documented examples involve attorneys or self-represented litigants, but UCLA law professors have warned that judges are likely to make more AI-related mistakes in the future. In recent months, the U.S. Senate examined apparent AI use by federal judges in Mississippi and New Jersey after rulings included significant factual errors.
Klapper, who previously worked as a law clerk for a federal appeals court and for the surveillance technology company Palantir, said courts need AI to reduce backlogs and operate more efficiently.
“Could we hire more staff?” he told CalMatters. “Maybe, but it won’t be enough to keep up with the exponential growth that is coming, or to adequately solve the current crisis. I think the only solution is to assign every judge and attorney their own AI assistant.”
Klapper said his goal is to combine the strengths of human judges with the capabilities of machines. He acknowledged that AI systems may contain bias but argued that models can be tested in ways humans cannot.
“I’m not saying all machines are free of bias,” he said. “I’m not even saying my machine is free of bias. What I’m saying is that we can test it, and people have tested it. That’s the advantage over humans.”
The generative AI policies for the Los Angeles and Riverside superior courts require disclosure only when a motion, ruling or other document is drafted entirely with generative AI.
Both courts declined to say whether litigants are being told that the tool is being tested in connection with their cases. A Los Angeles County Superior Court spokesperson told CalMatters the testing involves motions that have already been decided and is being done separately from active cases. However, the contract allows testing in live cases.
“Even with successful evaluation and extensive testing, the Court would still be several months, if not years, away from implementing this type of tool,” the spokesperson said.
The Los Angeles County contract permits the tool to be tested on two important types of criminal motions: suppression motions, which determine what evidence prosecutors may present at trial, and post-conviction relief motions, filed by people who have already been convicted and are seeking another chance at release.
Los Angeles County District Attorney Nathan Hochman said those provisions are his “biggest concern.” After reviewing the contract, he described the two categories as “incredibly important motions in the criminal justice system.”
“When you’re dealing with a person’s liberty — unlike the civil arena, which covers everything else — the stakes are of the highest importance,” Hochman said. “I don’t want to take the risk, especially in a criminal case, that AI gets it wrong. That someone’s constitutional rights are violated. That someone goes to prison wrongly or, on the other hand, that someone gets away with something.”
In Los Angeles, some judges learned more about the Learned Hand contract during a March presentation by Superior Court Judges Yvette Verastegui and Olivia Rosales, who oversee the criminal division. As part of an annual tour, the judges visit courthouses throughout the county to brief judges on court operations, caseloads and other issues. During a lunch, Verastegui and Rosales indicated the tool might someday be used to assist with Racial Justice Act petitions, according to a judge who attended.
The California Racial Justice Act allows people to challenge a criminal conviction or sentence if they believe racial bias played a role. People incarcerated in state prisons file petitions directly with the court. If a petition appears to have merit, the process can include appointment of counsel, legal briefing and evidentiary hearings before a judge decides whether to grant relief.
A tool such as Learned Hand could change that process. According to the judge who attended the lunch, Verastegui and Rosales said the software could generate proposed decisions for judges to consider when deciding whether to deny petitions or allow them to proceed to later stages.
“My concern, of course, is that courts will use that as an anchor and cling to that initial analysis,” the judge said. “It is an extremely dangerous path. Setting aside inaccuracy, which will be a major concern, it dehumanizes the entire process. It does not treat people as individuals with lived experiences. It essentially reimposes a one-size-fits-all model of justice.”
A second Los Angeles County Superior Court judge, who also spoke anonymously, recalled the presentation and said they would not trust or use the tool to summarize a Racial Justice Act petition.
AI can reproduce or intensify patterns found in the data used to build a model, including human bias. Large language models have shown racial and gender bias. Previous reviews of predictive policing technology used by the Los Angeles Police Department found racial bias, and an analysis of the COMPAS risk assessment algorithm found it was more likely to label Black people as at risk of reoffending after incarceration than white people with similar records.
Public defenders who spoke with CalMatters shared those concerns.
Elizabeth Lashley-Haynes, a deputy public defender in the Los Angeles County Public Defender’s Office, said it would be “extremely problematic and bordering on unethical” for a judge to use such a tool to review Racial Justice Act petitions, which she described as highly complex.
“They are unlike anything that has ever been done before in the legal system,” said Lashley-Haynes, who specializes in Racial Justice Act cases. “The words used in these cases, which have racial connotations or meanings, are far beyond the scope of anything artificial intelligence can do.”
In interviews with CalMatters, Klapper and Los Angeles County Superior Court Executive Officer David Slayton denied that the court has plans to use the tool for Racial Justice Act petitions. A court spokesperson later confirmed by email that the contract allows the tool to be used that way, but said that has not yet occurred.
Klapper said that if a Racial Justice Act module were developed, the tool would need to be evaluated for possible bias and built jointly with the court.
“The timing is very sensitive, right?” he said. “It’s a very delicate decision, I’m not going to lie. The stakes are high; I understand why people could be very concerned. Especially when it comes to criminal matters, I have even more hesitation and more caution than usual, because fundamental rights are at stake.”
In Los Angeles, six Superior Court judges and their research attorneys are using Learned Hand primarily for research, motion summaries and help drafting tentative rulings, according to Slayton. He said the tool will not be expanded beyond the civil division “until court leadership is comfortable.”
“The court is being very prudent and careful in its use of this type of technology,” Slayton said. “Until we evaluate it and determine that it is effective in those areas, we will not expand it to others.”
Slayton said the tool will be reviewed quarterly to determine how it might be used in the future, though he did not give specifics about the evaluation. A court spokesperson later told CalMatters by email that Learned Hand is judged by the same standards applied to law clerks and research attorneys: accurate legal research, sound analysis, neutral judicial writing and reliable work that supports judicial decision-making.
Los Angeles County Superior Court Judge Samantha Jessner, who chairs the court’s Judicial Technology Advisory Committee, said she did not know until recently that the tool might be used outside the civil division. She said judges are not involved in contract negotiations because of ethical restrictions.
“I think we have a duty and an obligation to explore whether artificial intelligence has a place in our work as a judiciary, and that is exactly what this pilot program is for,” Jessner said.
The Riverside County Superior Court signed its agreement with Learned Hand in February. Emails obtained by CalMatters show Klapper suggested to Riverside court executives Jason Galkin and Sarah Hodgson that the court begin with a common civil motion, then expand quickly once the tool was established. He suggested Hodgson prepare a list of motions and workflows that create the most difficulty, citing the Racial Justice Act as an example.
About two weeks later, Hodgson described the most labor-intensive motions as those “trying to drive us into retirement,” including discovery motions and attorney fee motions. For criminal matters, the court suggested Klapper focus on document-heavy issues, including death penalty habeas corpus petitions and probation revocations.
Since Riverside’s pilot began, seven civil and probate attorneys have had access to the tool. Galkin, the court’s executive officer, said the court is evaluating the product to understand what it can do. He said the tool is not being used to draft tentative rulings.
“We don’t even know whether expansion is likely, so there are no established criteria for what that expansion would look like or thresholds for it,” Galkin said. “Right now, the fundamental question is: Does this help staff and assist them in achieving their professional goals?”
As courts continue testing AI, attorneys such as Hochman say its use may be inevitable, but should be limited to routine, repetitive and lower-level tasks.
“It’s the analysis of the case itself, along with the conclusions that would be reached, that makes me very distrustful of AI right now,” Hochman said. “A large part of that is because I don’t know all the data it is using to make its decision. The one thing I am completely certain of is that AI did not go to law school.”
Original source: CalMatters













