California’s AI Court Clerk Reaches Into Criminal Cases

The Los Angeles County Superior Court signed a $314,000 contract with AI company Learned Hand in February, granting the startup access to the country’s largest trial court. Riverside County signed its own $10,000 agreement the same month. Both courts launched pilots without telling litigants their cases could be touched by the system.

Documents obtained by the nonprofit newsroom CalMatters show both contracts extend further than the courts have described publicly, authorizing AI assistance in criminal cases where defendants face incarceration or the reversal of a conviction. The distance between what the contracts permit and what the judges assigned to criminal courtrooms were told is the problem officials are now managing quietly.

Two Courts, Two Contracts, One Missing Disclosure Rule

The California Judicial Council required superior courts to adopt generative AI use policies before deploying such technology, and both pilots launched under those mandates. But the policies contain a significant gap: disclosure is required only if a motion, decision, or other document is written entirely by generative AI. Assisted drafting, research memos, and tentative ruling suggestions that Learned Hand actually produces fall below that threshold. A judge can receive an AI-generated research memo and a suggested tentative ruling, revise a paragraph, and sign it without any disclosure appearing in the public record.

Both courts declined to say whether plaintiffs are aware the tool is being tested on their cases. A spokesperson for Los Angeles County Superior Court said testing is conducted on motions that have already been decided, but the contract permits testing on live cases. The roughly 31-to-1 gap between the two contract values reflects different ambitions: Riverside County is explicitly exploring the technology, while the LA contract carries a detailed roadmap for expansion into criminal, family, and probate divisions.

Inside Learned Hand’s Architecture

Learned Hand combines language models from Anthropic, OpenAI, and Google to function as what the company calls an AI clerk. The tool can draft orders, produce research memos, and suggest tentative rulings. Shlomo Klapper, the company’s founder and chief executive officer, previously clerked for a federal appeals court and worked at Palantir, a data analytics and surveillance technology company. He has framed the product as a necessary response to court backlogs that hiring alone cannot close.

“I think the only solution is to give every single judge and staff attorney their own AI clerk,” Klapper told CalMatters. He acknowledged the tool may carry bias but argued testability is the relevant standard. “I’m not saying my machine isn’t even biased,” he said. “I’m saying we can test it and people have tested it. And that is the benefit over humans.” The company says it tests for bias and accuracy but has not published those results.

Roughly a dozen of the 51 superior courts that responded to CalMatters’ public records requests said they use AI tools from LexisNexis, Thomson Reuters, and Microsoft’s Copilot. The California Judicial Council’s court technology framework required each superior court to create a generative AI policy before deploying such tools – but those policies set the disclosure threshold that currently exempts assisted drafting from notification requirements. Learned Hand goes further than those off-the-shelf products, building its functions specifically around judicial drafting and offering courts an explicit expansion path into every case type, negotiated into each contract before any judge sees the tool run.

The Racial Justice Act Collision

California’s Racial Justice Act, passed by the legislature in 2020 as Assembly Bill 2542, gives incarcerated people the right to challenge a conviction or sentence they believe was infected by racial bias. Petitions arrive directly from prison. A valid petition is unlike any other criminal filing in the state system, requiring evidence that racial bias was a factor in the choice of charges, the plea, the verdict, or the sentence. The evidence is often woven into jury selection records, prosecutorial communications, and language patterns that legal specialists say require careful human interpretation.

The March Luncheon and Tentative AI Decisions

In March, Superior Court Judges Yvette Verastegui and Olivia Rosales, who lead the criminal branch of the Los Angeles court, presented to colleagues at a courthouse luncheon and said the tool could one day help process Racial Justice Act petitions. They described a workflow in which an AI-generated tentative decision would reach judges after a petition arrived from prison, either advancing or denying the case before a human judge made a final call on the record.

At least one judge in attendance was alarmed. “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,” that judge told CalMatters on condition of anonymity under judicial conduct rules. “It essentially reimposes a one-size-fits-all style of justice.” A second judge who attended said independently they would not trust the tool to summarize any Racial Justice Act petition.

Defenders Call It Bordering on Unethical

Elizabeth Lashley-Haynes, a deputy public defender at the Los Angeles County Public Defender’s Office who specializes in Racial Justice Act cases, called the application “highly problematic and bordering on unethical.” The petitions turn on contextual racial language she says no current AI can reliably parse. “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,” she told CalMatters.

Klapper and LA Superior Court Executive Officer David Slayton both denied the court has any current plans for this use. A court spokesperson confirmed the contract permits it regardless, describing the possibility as one that “has not commenced in any way.”

A Track Record That Gives Critics Ammunition

The case against AI in criminal proceedings is not theoretical. Large language models have a documented history of race and gender bias. An analysis of predictive policing technology used by the Los Angeles Police Department found racial disparities in its outputs, and a separate analysis of the COMPAS risk-scoring algorithm, widely used in criminal sentencing decisions, found it more likely to classify Black defendants as future crime risks than white defendants with similar records. AI systems can replicate or amplify those patterns from their training data.

• Nearly 90 California state or federal court cases have involved AI-related mistakes since August 2024, per researcher Damien Charlotin’s global tracking database
• $10,000 fine was issued to a Los Angeles-based attorney for citing cases that do not exist, described at the time as a historic sanction
• 4 Nevada County prosecution cases contained AI-linked errors, the Sacramento Bee reported this month
• 2 federal judges withdrew AI-tainted opinions in 2025 after admitting their offices used generative AI tools to draft rulings

Those federal judges, Henry T. Wingate of the U.S. District Court for the Southern District of Mississippi and Julien Xavier Neals of the U.S. District Court for the District of New Jersey, were pressed by Senate Judiciary Committee Chairman Chuck Grassley of Iowa to explain their withdrawn opinions. Both confirmed AI involvement: Wingate’s law clerk had used the tool Perplexity; Neals’s intern had used ChatGPT. Both opinions contained fabricated case citations and invented quotes attributed to litigants. Sterne Kessler’s 2025 review of AI hallucination sanctions across US courts found no consistent federal standard for addressing AI errors in judicial proceedings.

A 2025 MIT study found that language models from major companies can reduce critical thinking and brain activity in users, a compounding concern when judges receive AI-drafted memos they did not generate. UCLA Law School professors have predicted that more judges will make AI-linked errors as adoption spreads into courts at every level.

What the LA Contract Permits in Criminal Court

Los Angeles County District Attorney Nathan Hochman reviewed the contract and flagged two criminal motions it explicitly covers. The first is the motion to suppress, the pretrial hearing that determines what evidence prosecutors may present at trial. The second is the motion for post-conviction relief, filed by people already sentenced who seek to overturn their conviction or reduce their sentence. Hochman called these “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,” he told CalMatters. “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.” Internal emails obtained by CalMatters show Klapper pitching Riverside County executives on applying the tool to “things with the largest paper records” in criminal cases, including death penalty habeas corpus petitions and parole revocation hearings.

• Motion to suppress – LA contract, criminal division, explicitly covered
• Motion for post-conviction relief – LA contract, criminal division, explicitly covered
• Racial Justice Act petitions – permitted by contract language, not yet commenced
• Death penalty habeas corpus petitions – raised in Riverside internal emails
• Parole revocation hearings – raised in Riverside internal emails

An Oversight Framework Still Being Built

David Slayton, executive officer of the Los Angeles County Superior Court, confirmed the tool will be evaluated quarterly, though he offered no methodology specifics when initially asked. A court spokesperson later provided a standard: the tool is measured against expectations applied to human law clerks, including accurate legal research, sound analysis, and neutral writing. No independent auditor or third-party bias reviewer is specified in the criteria or in the contract.

Judge Samantha Jessner, who chairs the court’s Judicial Technology Advisory Committee, told CalMatters she became aware only recently that the tool could be extended beyond the civil division. Judges are excluded from contract negotiations due to ethical constraints on their role. Jessner said the court has a duty to explore AI’s potential, though she did not address how defendants whose motions are processed during the evaluation period would learn of the tool’s involvement, given the disclosure rules currently in place.

In Riverside, Jason Galkin, chief executive officer of the Riverside County Superior Court, kept his description measured. “We don’t even know if expansion is likely so there is no set criteria for what expansion might look like,” he said. His court’s seven users remain in civil and probate matters for now. Hochman’s line was more direct. “The only thing I’m 100% sure of is that AI didn’t go to law school,” he told CalMatters. “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.”

Quarterly evaluations for the Los Angeles civil pilot are underway. If the review satisfies court leadership, the contract permits expansion into criminal and family divisions, with no new notification requirement for defendants and no publication of bias test results required before that step. If it does not satisfy leadership, the criminal and Racial Justice Act provisions stay in the contract, dormant but intact, available for the next version of the conversation.