California Courts Test AI 'Clerk' for Judges

CalMatters reporting, reproduced by KQED and LAist, shows that two large California trial courts are piloting an AI tool sold by Learned Hand to draft orders and produce research memos for judges. CalMatters reports the Los Angeles County Superior Court opened a test in February under a roughly $314,000 contract that includes a roadmap to evaluate the tool in criminal, family and probate divisions. CalMatters also reports Riverside County has a $10,000 agreement to test the program primarily for civil and probate research memos. According to CalMatters, court officials declined to provide detailed criteria they will use to assess whether the technology can safely expand into higher-stakes criminal and family matters. CalMatters reports Learned Hand combines language models from Anthropic, OpenAI, and Google; the company says it tests for bias and accuracy but has not published results. One judge who spoke on condition of anonymity told CalMatters they were alarmed by comments from colleagues suggesting the technology could be used to evaluate certain appeals.

CalMatters reports that Learned Hand operates by leveraging third-party large language models from multiple vendors rather than a proprietary single model. The reporting does not include published evaluation metrics or the companys test results. Editorial analysis - technical context: When vendors stitch together models from multiple providers it can improve coverage for some tasks but also complicate provenance, version control, and explainability. Industry-pattern observations: Independent verification of model outputs, reproducible testing datasets, and chain-of-evidence logging are common requirements that courts and regulated institutions request before they accept AI outputs in adjudicative roles.

Deploying generative AI in courts intersects with procedural fairness, due process, and judicial ethics. Public reporting highlights three immediate concerns courts and practitioners commonly raise: transparency about whether an AI system influenced a judicial decision, auditability of the model inputs and outputs, and empirical validation that the system does not exacerbate bias in decisions that affect liberty. Reporting by CalMatters notes these concerns in the context of criminal appeals and the Racial Justice Act, where errors or opaque assistance could carry high human stakes.

• •Whether CalMatters or the courts publish vendor evaluation criteria or Learned Hands test results.
• •Any formal changes to court rules or local AI policies that govern use of automated drafting or research assistance.
• •Requests from defense counsel or appellants to disclose whether an AI system assisted in drafting orders, and any resulting legal challenges or guidance from judicial conduct commissions.

Editorial analysis: For practitioners and vendors, this pilot underscores the need for documented evaluation frameworks, versioned model provenance, and auditable output trails when AI supports decisions in regulated, high-stakes domains. Observers will also watch whether courts require vendor-published validation or third-party audits before allowing AI assistance in criminal or family matters.