California Judges Are Testing a New AI Clerk. You Won’t Know if It’s Looking at Your Case

Learned Hand uses a combination of language models from Anthropic, OpenAI and Google to act as an AI clerk for judges. The company says it tests for bias and accuracy, but it has not yet published results.

In Riverside County, which has a $10,000 agreement with the company to test the program, civil and probate attorneys are primarily using the tool to draft research memos that help judges reach their decisions. It’s typical for research attorneys to assist judges as they review cases.

Los Angeles County Superior Court has a roughly $314,000 contract that includes a roadmap to test the tool’s use in criminal, family and probate divisions.

Officials would not describe in detail to CalMatters the criteria they’re using to evaluate whether use of the tool can safely expand to criminal and family courts, where the stakes are often much higher than in civil cases.

One judge who spoke to CalMatters on condition of anonymity due to judicial rules of conduct was alarmed when their colleagues at a recent luncheon said the technology could be used one day to evaluate appeals from people who believe their conviction or sentence was tainted by racial bias. California courts are handling a wave of those claims after lawmakers passed the Racial Justice Act in 2020.

“I think it is outrageous,” said the Los Angeles County Superior Court judge. “AI cannot and never will be able to replace human judgment in evaluating complex social dynamics. Ultimately, that will erode the public’s confidence in the competence and fairness of the judiciary.”

A majority of California’s superior courts now have generative AI use policies, according to documents obtained by CalMatters via public records requests, which they were required to create by the state Judicial Council before using the technology. Roughly a dozen of the 51 courts that have responded to CalMatters’ requests said they are using AI-powered tools from LexisNexis, Thomson Reuters, and Microsoft’s Copilot.

Use of AI in courts has been controversial because of the propensity of AI models to cite falsehoods and to produce sycophantic text. Models from major companies like Google and Anthropic can reduce critical thinking and brain activity, according to a 2025 MIT study.

Language model hallucinations have already made it into the judicial system. Researcher Damien Charlotin has documented hundreds of instances of litigants, lawyers, and judges making mistakes when using AI to do their jobs including nearly 90 cases in state or federal courts based in California since August 2024.

Last fall, a Los Angeles-based lawyer received a historic $10,000 fine for citing cases that don’t exist, and earlier this month the Sacramento Bee reported that use of AI led to errors in four cases handled by prosecutors in Nevada County. Most of these cases involve lawyers or people who are representing themselves in court, but UCLA Law School professors predict that more judges will make AI-fueled mistakes in the future. In recent months, the U.S. Senate investigated federal judges in Mississippi and New Jersey for drafting decisions with generative AI that had serious factual errors.

Klapper, who previously worked as a clerk for a federal appeals court and for surveillance technology company Palantir, said the judiciary needs AI in order to reduce backlogs and increase efficiency.

“Could we hire more people?” he told CalMatters. “Maybe, but it’s not going to keep pace with the exponential increase that’s coming, nor is it going to be able to adequately solve the crisis of today. I think the only solution is to give every single judge and staff attorney their own AI clerk.”

Klapper said he’s aiming to combine the best parts of what human judges can do with the best parts of what machines bring to bear.

“I’m not saying all machines aren’t biased,” he said. “I’m not saying my machine isn’t even biased. I’m saying we can test it and people have tested it. And that is the benefit over humans.”

Generative AI use policies for the Los Angeles and Riverside County superior courts only require disclosure if a motion, decision, or other document is written entirely with generative AI.

Both courts refused to say whether plaintiffs are aware that the tool is being tested on their cases. In a statement to CalMatters, a spokesperson for the Los Angeles County Superior Court said testing is done on motions that have already been decided, separate from live case environments. However, the contract allows for testing on live cases.

“It is important to note that even with successful evaluation and thorough testing, the Court remains several months, if not years, away from implementing this type of tool,” said the spokesperson.

The contract allows the tool to be used for two critical motions in the criminal division: A motion to suppress, which is designed to determine what type of evidence the prosecution is allowed to present at trial, and motions for post conviction relief, which are filed by people who have already been convicted and want another shot at freedom.

That’s the “greatest concern” for Los Angeles County District Attorney Nathan Hochman. When he reviewed the contract, he referred to the motions as “two incredibly important motions in the criminal justice system.”

“When you’re dealing with someone’s liberty — as opposed to in the civil setting, which is everything other than liberty — the stakes couldn’t be higher,” said Hochman. “I don’t want to take the chance, particularly in a criminal case, that AI happens to get it wrong. And now someone’s constitutional rights have been infringed. Someone has gone to prison who shouldn’t have, or on the flip side, that somehow someone gets off.”

‘An extremely perilous road’

In Los Angeles, some judges first heard about the new Learned Hand contract during a March presentation by Superior Court Judges Yvette Verastegui and Olivia Rosales. They lead the criminal branch and visit courthouses throughout the county as part of an annual roadshow, where they update judges on court operations, discuss workload and field questions. During a luncheon, Verastegui and Rosales said the tool could be used to assist with Racial Justice Act petitions in the future.

California’s Racial Justice Act allows people to challenge a criminal conviction or sentence that they believe was based upon racial bias. Petitions are filed directly to the court from people in state prison. If a case is found to have merit, the process includes appointing legal counsel, filing briefs and setting evidentiary hearings before a judge would decide whether to grant the petition.

That process could look different with a tool like Learned Hand. Verastegui and Rosales explained that, following an incarcerated person’s petition, the tool could generate tentative decisions for judges to consider in denying or advancing cases to the next stages, according to one judge who attended the luncheon.

“The concern, of course, that I have is that the courts will utilize that as a reference point and then get stuck to that initial analysis,” said the judge. “It’s an extremely perilous road to go down. Putting aside the inaccuracy, which will be a significant concern, it dehumanizes the whole process. It does not treat people as individuals with lived experiences. It essentially reimposes a one-size-fits-all style of justice.”

A second Los Angeles Superior Court judge who spoke with CalMatters on the condition of anonymity remembered the presentation and said they would not trust nor use the tool to summarize a Racial Justice Act petition.

AI can replicate or intensify patterns contained in the data used to make a model, including human biases. Large language models have a history of demonstrating race and gender bias, an analysis of predictive policing tech used by LAPD found racial bias, and an analysis of the risk assessment algorithm COMPAS found that it is more likely to label Black people as at risk of committing crimes after incarceration than white people with a similar record.

Public defenders who spoke with CalMatters echoed those concerns.

Elizabeth Lashley-Haynes, a deputy public defender at the Los Angeles County Public Defender’s Office, said it would be “highly problematic and bordering on unethical” for a judge to use the tool to review Racial Justice Act petitions, which she described as “incredibly nuanced.”

“They’re like nothing else in the legal system that has ever really been done,” said Lashley-Haynes, who specializes in Racial Justice Act cases. “Words that are used in these cases that have racial undertones or racial meanings are way beyond the realm of anything that artificial intelligence could do.”

In interviews with CalMatters, Klapper and Los Angeles County Superior Court Executive Officer, David Slayton, denied that the court has any plans to use the tool for Racial Justice Act petitions. A spokesperson for the Los Angeles Superior Court later confirmed in an email to CalMatters that the contract permits the tool to be used in such a way “but that possibility has not commenced in any way.”

Klapper said if they were to build out a Racial Justice Act module, the tool would need to be evaluated for bias and co-developed with the court.

“The timing very fortuitous, right?” he said. “It’s a very fraught decision, I’m not going to lie…extremely high stakes — a scenario where I understand people might be very concerned. Especially with criminal, I have even more hesitancy, even more guardrails than normal about, because there are liberty interests at stake.”

Extending beyond civil cases

In Los Angeles, six superior court judges and their research attorneys are primarily using the Learned Hand tool to conduct research, summarize motions and assist in drafting tentative rulings, according to Slayton. He says the tool won’t move beyond the civil division “until the court leadership is comfortable.”

“The court is being very deliberate and careful about how we use technology like this,” he said. “So until we evaluate it and determine that it is effective in those areas, we will not extend it to other areas.”

The tool will be evaluated on a quarterly basis to determine its future application, Slayton said, but he did not specify what kind of evaluation that entails. In an email to CalMatters, a spokesperson later said that Learned Hand is evaluated “against the same substantive expectations applied to law clerks and research attorneys: accurate legal research, sound analysis, neutral and judge-ready writing, and reliable work product that supports judicial decision-making.”

Los Angeles County Superior Court Judge Samantha Jessner, who chairs the Judicial Technology Advisory Committee, said she was unaware of the possibility that the tool could eventually be used outside of the civil division until recently. Judges are not privy to contract negotiations due to certain ethical limitations, she said.

“I think we have a duty and obligation to explore whether or not there is a place for artificial intelligence in what we do as a judicial branch and that’s exactly what this pilot is intended to afford us the opportunity to do,” said Jessner.

Riverside County Superior Court signed an agreement with Learned Hand in February. In emails obtained by CalMatters, Klapper proposed to two Riverside County Superior Court executives, Jason Galkin and Sarah Hodgson, that the court use the tool for a common civil court motion and “then expand quickly once we earn our stripes.” He suggested that Hodgson assemble a list of motions and workflows “that generate the most pain,” citing examples that included the Racial Justice Act.

Roughly two weeks later, Hodgson described the most laborious motions “that want to drive us into retirement,” including discovery motions and attorney fee motions. For criminal cases, the court suggested that Klapper focus on “things with the largest paper records,” citing death penalty habeas petitions and parole revocation.

Since the pilot started, seven civil and probate attorneys have been granted access to the tool. Galkin, the chief executive officer of the Riverside County Superior Court, said they are “kicking the tires on the product” to see what tasks it can do. The tool is not being used to draft tentative rulings, he said.

“We don’t even know if expansion is likely so there is no set criteria for what expansion might look like or thresholds for that because right now, the core question is: Does this help staff and does it advance what they’re trying to do in their roles?” said Galkin.

As testing is underway, attorneys like Hochman say that use of AI is inevitable, but would be better suited for low-level, repetitive and routine tasks.

“It’s the analysis of the case itself, coupled with the conclusions that will be reached, that I’m very hesitant to trust AI at this point — in large part, because I don’t know all of the inputs that AI is using to make its decision. The only thing I’m 100% sure of is that AI didn’t go to law school,” said Hochman.

Cayla Mihalovich is a California Local News fellow.

This article was originally published on CalMatters and was republished under the Creative Commons Attribution-NonCommercial-NoDerivatives license.