Live AI Interpreting for Public Services: The Evidence Stops Where Stakes Rise

The problem

California's courts listed 1,856 certified and registered court interpreters as of March 2025, covering 114 spoken languages and American Sign Language, in a state where more than 200 languages are spoken. The courts' own study refers to interpreter shortages in some courts and an aging workforce (court-reported, June 2025)¹. Spanish accounts for about 88.0% of all interpretations over the study period, so every other language shares a small remainder of the demand¹.

The state's hiring pilot shows the direction of travel. Court interpreter employees fell from 799 in 2021-22 to 717 in 2023-24, and the pilot received 1,114 applications by 1 August 2024, 90% of them for Spanish (court-reported)². In 2022-23 there were 377,070 interpreted Spanish events against 1,310 Spanish interpreters (court-reported)².

Outside courts, a survey by an interpreting-software vendor found that 50% of its 370-plus respondents said their organisation had met a person with limited English without an interpreter when one was needed (vendor-reported, self-reported, non-random sample, date not shown)⁷. Trade coverage of a 2026 review reports an estimated 25.7 million Americans with a non-English language preference, and written discharge instructions that can take one to five days to translate at large academic centres (secondhand, as reported)⁸.

The evidence comes from the United States and Canada. No European or Asian data on the problem was read, and the private sector appears only through the vendor survey.

What already works

Speech recognition and machine translation already run in consumer tools. The one peer-reviewed performance test read for this paper used Google Translate's conversation mode (version 6.55.0) between February and April 2023³. It was a simulated, scripted low-acuity paediatric consultation (a fever with a respiratory infection) in five languages: Spanish, French, Urdu, Arabic and Mandarin. An interpreter played the caregiver and a clinician played the physician, with one test of about 180 minutes per language and no real patients³.

The final share of correctly interpreted paragraphs, counting paragraphs corrected within up to two retries, ranged from 83.5% (Urdu) to 95.4% (French). First-attempt accuracy ranged from 76.1% (Spanish) to 91.7% (Arabic). Paragraphs still misinterpreted after the retries ranged from 0.9% (Arabic) to 6.4% (Urdu) (measured in a simulated test, one run per language)³. In plain terms, most of each conversation got through, but in the weakest language roughly one paragraph in sixteen stayed wrong.

The errors were mostly wrong words rather than omissions. "Congestion" came out as "traffic jam" in Arabic, Urdu and Mandarin, and "2 or 3 extra" became "between one and 2000 diapers" in Urdu³. The authors concluded that the tool can improve communication in non-critical consultations when no interpreter is available, but requires careful execution, and noted that one test per language limits reproducibility³.

Human interpreters are not error-free either. The same paper cites earlier research reporting an average of 31 mistakes in human-interpreted medical visits (cited secondhand; the original was not read)³. For written text, trade coverage of a 2026 review reports clinically significant error rates for Haitian Creole of 8.3% for professional human translations, 23.3% for Google Translate and 33.3% for ChatGPT (secondhand; the underlying paper was not read; written discharge text, not speech)⁸.

Remote human interpreting is also established. In California courts, 93% of interpretations were in person, while remote video peaked at 48,892 and telephone at 16,765 interpretations in 2020-21. The study recommends expanding remote and telephonic interpretation to close coverage gaps, especially for less common languages and rural courts¹.

What does not exist. No trial of live machine interpreting with real patients, defendants or residents was found. The authors of the one study read say further testing in real-life contexts is needed³. No source read showed a system that reliably signals when its own translation is wrong.

A first legislative test in Wisconsin

Wisconsin Senate Bill 357 (2025) is described in its official title as relating to "use of artificial intelligence or other machine-assisted translation tools in lieu of appointing English language interpreters", alongside making English the official state language⁴. It was introduced on 27 June 2025 by Senators Jacque and Nass and referred to a Senate committee. The latest action on the Legislature's page is a fiscal estimate received on 15 July 2025, with no recorded hearing or vote (official record; the page's cache date is unknown)⁴.

A local news report dated 5 March 2026, carrying an "AI Assisted" label, describes amendments that would bar AI interpretation in proceedings involving an accused violent crime and allow virtual interpreters at trial only if both sides agree, and says the package had seen floor action. None of this appears on the official page, so it is treated here as unconfirmed⁶. The report says backers argue that rural counties struggle to find interpreters and quotes the sponsor as saying AI could spare local governments burdensome expenses⁶.

The National Association of Judiciary Interpreters and Translators, which says it represents over 1,200 court interpreters, translators and others, wrote on 15 July 2025 advising withdrawal of the bill. It concedes that technology "could assist courts and human interpreters in specific contexts", but objects that the bill names no approved tools, sets no accuracy validation, names no party accountable for AI output, and raises error risk for languages of lesser diffusion (advocacy by a professional association with an interest in the outcome)⁵.

The possibility

One way this could work is to tier interpreting by stakes. For routine contacts, such as opening hours, form help or benefits queries at a council counter, a machine could interpret alone, with a notice and a one-touch option to ask for a person. For consultations that matter but are not irreversible, such as a non-urgent clinic visit or a licensing hearing, a machine could carry the conversation while a human interpreter stays on remote standby to listen in, correct or take over, with key points confirmed in writing in the person's language. For criminal trials, consent to surgery, child protection or asylum, a qualified human interpreter would always be used, and a machine could at most assist that interpreter.

The role of AI. Four parts would depend on AI: speech recognition, translation, synthetic voice, and language identification. The fifth part is the hardest and would also depend on AI: judging when a given translation is unreliable and routing the conversation to a human. Everything else is ordinary systems, logistics, money or policy: the video or phone link, the booking and pay of human interpreters on call, consent forms, recording and retention rules, procurement, staff training and privacy law. If the evidence points the right way, much of the gain in access might come from cheap, well-organised remote human interpreting, with AI supplying a first tier and filling gaps; the California study already recommends remote and telephonic expansion on its own¹.

If the machine tier lowers the cost per minute, the savings could fund human interpreters for the high-stakes tier. If the savings were instead taken as a budget cut, the same arrangement could quietly replace the humans it was meant to protect.

What would have to be true

1. Accuracy measured where it matters. Accuracy would need to be measured per language, dialect and setting, with real participants. The one test read covered five languages in a simulated low-acuity scenario and ranged up to 6.4% of paragraphs still wrong in Urdu, so a single average would hide the weakest language³.
2. The system knows when it is wrong. It would need to detect its own uncertainty and hand over to a person. This is an AI capability condition, and no source read demonstrated it for speech interpreting.
3. A right to a human. People would need a real right to ask for one. The BMC paper notes that the Canadian Charter gives a right to an interpreter in legal proceedings but not in healthcare encounters, and that few legal routes exist for patients to assert one³. Other jurisdictions were not checked against official texts.
4. Data handling that protects speech. The tested tool processes speech on remote servers, which the authors identify as a privacy and security risk. They recommend informed consent, avoiding personal identifiers and considering privately hosted versions³.
5. A human workforce that survives. Employee interpreters in California courts fell from 799 to 717 between 2021-22 and 2023-24². If a machine tier reduced training and hiring further, the high-stakes tier could have too few people left to staff it.
6. Someone accountable. A named party would need to be responsible for machine output; the judiciary interpreters' association says the Wisconsin bill names none⁵.
7. Fair performance for everyone. The authors note that performance is best with standard pronunciation, which could disadvantage dialect speakers, young children, people with limited literacy and neurodivergent people³.

What could stop it

What the evidence shows

• Only one peer-reviewed test was read, it was simulated, and it ran once per language; the authors say reproducibility is limited³.
• Tested errors included wrong words with potential consequences, such as "2 or 3 extra" becoming "between one and 2000 diapers"³.
• The professional association for judiciary interpreters opposes Wisconsin's bill on grounds of unspecified tools, no accuracy validation, hallucinations and no accountability (advocacy)⁵.
• Opposition from the State Bar of Wisconsin, Disability Rights Wisconsin and the ACLU of Wisconsin is reported by a news source with an AI-assisted label and is unconfirmed⁶.
• In the vendor's survey, 16.8% of respondents used AI interpreting, 15.9% were evaluating it and 67.3% were not using it (vendor-reported, self-reported, non-random sample)⁷.
• The bill itself shows no recorded action beyond a fiscal estimate in July 2025 on the official page⁴.

Reasoning

The challenge question is what would make this less likely, less useful or more harmful than it first appears. The first answer is about failure. If a system cannot tell when it is out of its depth, the arrangement fails exactly when the stakes are highest.

The second is about who bears the risk. The person being interpreted usually cannot check the output in the language they are being translated into, so a wrong "yes" or a wrong dose could pass unnoticed by the one person most affected. A notice and an opt-out would help only if the person knew a mistake had occurred.

The third is about where the value lies. In the one test read, Urdu scored lowest³, and Spanish, the language with the deepest human supply, is the one where human interpreters are most available¹². If low-resource languages generally perform worse, the machine tier might add the most value in precisely the languages where it is least accurate, which would call for the most cautious human backup there.

The fourth is about simpler options. Pre-translated forms, bilingual staff, plain-language redesign and remote human interpreters on call could cover part of the need without AI. The evidence read does not show how much of the gap such measures could close, so the added value of AI would have to be demonstrated against them, not against doing nothing.

Signs to watch

• Any recorded hearing, amendment or vote on Wisconsin SB 357 on the Legislature's bill page⁴.
• Results of California's five-year Court Interpreter Workforce Pilot, launched in 2024 against a baseline of 717 employee interpreters in 2023-24².
• A published study of live AI interpreting with real patients or hearings, which the authors of the simulated test call for³.
• A court or health authority rule naming which interpreting tools may be used, for which proceedings, and who is accountable for output.
• The next Judicial Council of California language need and interpreter use study, which would show whether interpreter numbers and remote use have changed¹.