Domain Atlas / Caseworker documentation & copilots
Learned Hand AI clerk pilot (LA and Riverside courts)
In February 2026 the Superior Court of Los Angeles County, the largest trial court in the United States, began a pilot of the Learned Hand AI drafting workbench with six civil-division judges and their research attorneys under a contract of about $314,000 running into early 2027, and the Superior Court of Riverside County gave seven civil and probate research attorneys access under a separate $10,000 agreement used for research memos; the tool ingests case filings, synthesizes applicable law, and drafts proposed orders in the individual judge's own writing style. Under California Judicial Council Rule 10.430 (effective September 1, 2025, the first statewide court generative-AI framework in the nation), disclosure is required only when a document consists entirely of generative-AI output, and the rule reaches judicial officers only for tasks outside their adjudicative role, so neither court is obligated to tell litigants when AI assisted with an order or memo in their case; both courts declined to confirm whether litigants whose cases are used in testing are informed.[3]
What happened
On March 18, 2026, the vendor Learned Hand — which describes itself as the only AI company built exclusively for the judiciary — and the Superior Court of Los Angeles County announced a partnership giving a select group of judicial officers access to an AI workbench for case review, summarization, research, analysis, and drafting assistance. The Los Angeles court is the largest trial court in the United States: roughly 1.2 million cases filed annually, nearly 600 judicial officers, 36 courthouses, serving more than 10 million residents. The pilot had begun a month earlier, in February 2026, with six civil-division judges and their research attorneys under a contract of about $314,000 running into early 2027, applied to civil motions including motions for summary judgment and approval of class-action settlements. Separately, the Superior Court of Riverside County signed a $10,000 agreement in February 2026 giving seven civil and probate research attorneys access, used for drafting research memos rather than tentative rulings.
The tool ingests case filings, organizes the record, synthesizes applicable law, and drafts proposed orders using samples of the individual judge's own prior writing for style. The vendor states that every output is hyperlinked to the source material within the case file, and reporting describes a fact-checking process the company brands "Deep Verify" that interrogates every sentence of a generated order; the vendor also claims multiple verification passes and adoption by the Michigan Supreme Court and trial courts in ten states. Those are vendor statements, and no independent evaluation of the tool has been published. The only error-correction safeguard reported is the judge's own review: officers in the pilot are required to review and edit the draft before adopting tentative rulings, and a court spokesman said the assistance does not supplant the judicial officer's independent role. No external audit, query logging, or benchmarking regime was reported, and legal analysis coverage drew a contrast with Michigan's stronger governance, which it described as including external audits, query logging, and benchmarking.
Neither court is obligated to tell litigants when AI assisted with an order or a research memo in their case. Under the California Judicial Council's Rule 10.430 — adopted effective September 1, 2025, the first statewide court generative-AI framework in the nation, which required courts that do not ban the technology to adopt use policies by December 15, 2025 — disclosure is mandated only when a document consists entirely of generative-AI output, and the rule applies to judicial officers only for tasks outside their adjudicative role. Partial assistance on an adjudicative task, which is what this tool provides, therefore triggers no disclosure. Both courts declined to confirm whether litigants whose cases are used in testing are informed; the Los Angeles court said testing occurs on already-decided motions, though the contracts permit use on live cases, so whether any live case has been decided with an AI draft is not established in the public record.
Backlog pressure is the explicit adoption driver. Coverage describes a workload crisis and a "paper blizzard" driven partly by rising numbers of self-represented litigants; one measured surge is Los Angeles employment-litigation filings rising 49% in a year, from 4,100 to 6,400, per a February 2026 Fisher Phillips report — a subcategory of the docket, not the whole. The vendor's chief executive, Shlomo Klapper, whose company was founded in 2024, said courts are under tremendous strain, that their caseloads rise but no help is coming, and that the only solution is to give every single judge and staff attorney their own AI clerk. Critics raise the opposite concern. The Los Angeles District Attorney, Nathan Hochman, warned that an AI-generated draft ruling could greatly influence what the judge's position should be — the draft arrives before the judge forms an independent view — and, on the roadmap's contemplated expansion into criminal cases, that when you are dealing with someone's liberty the stakes could not be higher. That roadmap includes the criminal, family, and probate divisions, specifically motions to suppress and post-conviction relief motions; one judge told CalMatters that colleagues had discussed using the tool to evaluate California Racial Justice Act petitions, a use a public defender, Elizabeth Lashley-Haynes, called highly problematic and bordering on unethical. As environmental context for hallucination risk, the researcher Damien Charlotin has documented nearly 90 California state and federal court cases involving AI-fabricated content in filings since August 2024 — these involve litigant and attorney filings, not the Learned Hand tool itself. Los Angeles County plans quarterly evaluations of the pilot, with criteria unspecified beyond the standards applied to human law clerks; as of mid-2026 no evaluation results had been published, and court officials said they were several months, if not years, from broader implementation.
The sociotechnical reading
The Atlas already carries copilots whose entire safety case is a human review, and it is worth being precise about why this one is different. Magic Notes is a vertical documentation tool whose one review gate drifts toward approval under load; the Nava and Imagine LA benefits navigators are verify-before-use tools where a caseworker reads a cited answer before relaying it; the cross-government productivity layer's finding is a measured trade. What those share, and what this case removes, is a downstream channel. A benefits caseworker who relays a wrong answer can be corrected by an appeal, a supervisor, or an audit; a documentation error can be caught when the next worker reads the file. The Learned Hand pilot sits at the adjudication node itself — the single point where the system's output becomes a binding court order — and at that node the operator is the terminal authority. There is no supervisor above the bench, and in this deployment no external audit, no query logging, and no benchmarking were reported. The judge's own review is not the first line of defense with backstops behind it; it is the only line.
That is the distinct lesson this case adds to the collection: a single human review is a genuine safeguard only when something downstream can catch what the review misses, and here two structural facts make sure nothing can. The first is the missing disclosure edge. Because California's rule is triggered only by a document written entirely by AI, partial assistance on an adjudicative task is never disclosed, so the litigant who would have standing to challenge an AI-induced error is never told to look for one. The second is style imitation. The copilot drafts in the individual judge's own voice and reads the judge's prior rulings back to do it, so an AI-authored line is hard to distinguish forensically from the judge's own prose — even a careful reader downstream cannot tell which lines to question. Put together, a review at a terminal adjudication node with no disclosure and no forensic signature is not an error-correction channel; it is the absence of one, dressed as sufficiency.
The system map makes the trap legible. The load-bearing dynamic is a feedback loop: backlog pressure is the stated reason to adopt the tool, and the tool's job is to absorb backlog by drafting more — but the review the whole safeguard depends on runs on exactly the time the backlog is consuming, so the force that recruits the tool erodes its only check. A second loop forms on the input side, where AI-assisted filings from self-represented litigants swell the docket that drives further adoption. And the style-imitation self-loop closes the record: an AI-assisted draft the judge adopts becomes a binding order, feeds the prior rulings the copilot reads back to imitate the judge, and so shapes the next draft, all of it unmarked. The productive levers follow directly and are exactly what the deployment lacks — an independent check between an AI draft and the binding record, a gate on what a machine writes into that record, provenance so an AI-touched line carries its origin, and protected verification capacity that holds when the caseload does not. None of this asserts that the tool has produced a wrong order; no error rate exists, the anchoring concern is a contested critique rather than a measured effect, and the honest reading is structural. The point is that at the adjudication node the usual sentence "a human reviews it" stops being a safeguard and becomes a description of where the safeguards end.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.