California’s Judicial Council has canceled a plan to roll out AI translation at court counters statewide, saying the service must undergo a review of its accuracy and efficiency.
San Diego Superior Court, one of more than 30 courts that piloted the voice-to-text tool, used it about once a month in its Civil Business Office from 2022 until this spring, but clerks preferred a telephone translation service, officials said.
The court collected informal feedback on the AI tool from staff, said Emily Cox, a spokesperson for San Diego Superior Court. Cox said staff reported that “it was difficult for the application to capture information from people who spoke quickly,” and that the tool occasionally required software updates that delayed service. In addition, she said the tool was unavailable in Mandarin and Haitian.
While the pilot has ended, staff at the San Diego court continues to use a telephone translation service offered through a private vendor, United Language Group, which Cox said uses human interpreters. The service was available throughout the pilot, alongside the tool.
The pilot program began in 2021 and ended in April 2026, according to public records obtained by legal aid groups. Before launching the pilot, a work group of the Judicial Council explored available technologies, including Google, Amazon, IBM and Microsoft. Google had the highest average score in tests the council ran for accuracy and responsiveness.
The Judicial Council, which oversees policymaking for California courts, released a request for proposals in late August seeking vendors to roll out AI translation statewide at public-facing service points such as clerk’s counters, with a contract starting Dec. 1. Vendors submitted proposals in late September, and the council was scheduled to announce its intent to award a contract Oct. 26. Instead, it published a notice of cancellation on Oct. 15.
A spokesperson for the Judicial Council said in an email that the request for proposals was cancelled, citing evolving technology as the reason. The council plans to seek an internal review relating to “accuracy, efficiency, and effectiveness” before considering any new voice-to-text options.
The council has said, however, that it conducted a preliminary evaluation of the pilot but has not made the results public.
The cancellation notice came the day before the council’s Friday, Oct. 16, meeting, where advocates raised concerns about the rollout after reviewing transcripts from the pilot project.
Legal aid groups, including the Legal Aid Foundation of Los Angeles and California Rural Legal Assistance Inc., obtained hundreds of pages of transcripts of the program’s translations through public records requests. They said professional translators who reviewed the transcripts found multiple errors across different languages.
The transcripts do not identify which courts they came from, so the errors cannot be tied to San Diego or any other individual court.
Joann Lee, special counsel at the Legal Aid Foundation of Los Angeles, said some of the transcripts showed critical information had been wrongly translated.
“That could really have harmful effects on [court users’] livelihoods — like paying fines, like their ability to get restraining orders, even respond to criminal proceedings,” she said.
In one exchange with an Arabic-speaking court user seeking help with a restraining order, the transcripts show a court staff member had told a court user to fill out a sheet with the name of the one “who’s getting the restraining order, phone number and address.” According to the advocates’ review, the AI tool translated the statement as instructions to fill out the form with the name of the person “who receives the phone number, and address of the restraining order.”
Other problems the agencies found in the transcripts include:
- Incorrect translations, including DUI being translated as IUD.
- Misinterpretations of legal terminology, including translating “no contest” as “no competition.”
- An inability to clarify the subject, as in translating “fear that he/she/they could do something to me,” as “fear that you do something to me.”
These errors, advocates say, could undermine court users’ ability to exercise their legal rights and obtain critical information.
The advocates also reviewed survey results, obtained through public records requests, based on court staff feedback about the tool.
The staff flagged similar issues to what the advocates noted in their analysis of the transcripts. They include allegations that the tool had difficulty in hearing users correctly and that it picked up surrounding noise, impacting its ability to translate properly. Court staff also said that the tool did not:
- Translate complex legal and technical terms correctly
- Translate correctly when the court user spoke in improper Spanish
- Translate consistently across languages. It handled some better than others.
Many of these concerns are issues that could be discerned by human certified interpreters. Those interpreters might not always be 100% accurate, Lee said, but the errors humans make are different from those a machine would make.
“Maybe they [human interpreters] are leaving out some details, but it’s not completely the wrong thing,” Lee said. “You might be giving really incorrect and harmful information through the machine.”
Or, as Norma Ventura, program director at California Rural Legal Assistance Inc., said, a literal translation “might be correct but contextually is inappropriate and renders the whole sentence incorrect.”
Marine Carpuat, a professor of computer science at the University of Maryland who studies machine translation, said that every translation system makes mistakes and so potential errors should be taken into account.
“I think there’s real potential for these systems, but we need to acknowledge that they’re imperfect and design for it,” Carpuat said.
Advocates remain concerned about future efforts to use AI for translation without guardrails, including qualified human review.
“It’s more important than ever to raise awareness regarding the harms AI technologies have on vulnerable and linguistically marginalized communities, to prevent proposals from being solicited in the future,” Lee said in an email.
