The AI translation assistant is, in fact, an existing, off-the-shelf AI tool powered by a large language model (LLM), such as ChatGPT or Claude, that has been trained on UI data.
More specifically, the materials — a 42-page slideshow with step-by-step instructions — are geared toward unemployment insurance (UI) systems.
New Jersey’s own use of AI assistants for translation has been in the works since 2018, when Gov. Phil Murphy took office and considered how to improve access to benefits, including unemployment benefits. According to USDR, unemployment applicants who do not speak English as their primary language are 50% less likely than English speakers to receive UI benefits.
The Covid-19 pandemic, which sparked a deluge of applications emphasized the need for improved language access. New Jersey lawmakers also approved a legal mandate in January 2023 requiring the state, including the agency responsible for UI, to improve language access through interpreting and translation.
One specific goal of New Jersey’s Department of Labor was expanding the state’s glossary of terms on the unemployment website. The Employment Insurance Modernization Project turned to New Jersey’s bilingual UI call center agents for their specialized knowledge of technical terms in Spanish.
In 2023, New Jersey trained the AI assistant, using it to draft a new glossary of Spanish unemployment terms vetted by bilingual call center agents. DOL published the glossary just as generative AI (e.g., ChatGPT) was first commercialized.
Staff have had access to the assistant since April 2024, with apparently positive results: According to New Jersey officials, the AI assistant can translate text three times faster than a human, with a 90% accuracy rating. USDR reports that their methods have reduced application times by 80% and achieved parity in application completion rates between English and Spanish speakers.
2024 Slator Pro Guide: Translation AI
The 2024 Slator Pro Guide presents 20 new and impactful ways that LLMs can be used to enhance translation workflows.
No Coding or Technical Expertise Required
USDR is a non-partisan, “tech-neutral” nonprofit with the goal of helping the government better meet the public’s needs. The organization has established relationships with 400 government agencies.
The document shared on December 5, 2024, describes the training materials as “no-code language resources for plain English and Spanish,” specifically for UI translations.
The kit includes research and testing results to validate the approach; step-by-step instructions to create an AI translation assistant; a translation evaluator to measure output quality; custom prompts; and several example workflows and use cases.
The process starts by translating the original English language UI content, including complex UI technology, into plain English that laypeople can understand. The agency then conducts UX research on the plain English and modifies it based on feedback. The AI assistant then translates the plain English text into plain Spanish, followed by UX research, feedback, and modifications for the Spanish version.
“While your AI assistant will create a strong initial translation, having a native Spanish-speaker review and edit the text ensures accuracy and natural language flow,” the materials advise. “If bilingual staff aren’t available, you can use your AI assistant to generate alternative versions and measure them using our translation evaluator.”
Slator Translation as a Feature (TaaF) Report
The Slator Translation as a Feature (TaaF) Report is a vital and concise guide on how AI translation is becoming an integral feature in enterprise technology.
Finally, the validated translations can be made publicly available. (The final translations can also go through the process again as part of future training for the AI assistant.)
USDR does warn of some potential accuracy issues: incorrect technical terms; inconsistent terminology; translations that do not fully match the original meaning; and missing or added (i.e., hallucinated) information.
But, the materials note, “[a]gencies can implement this approach immediately using off-the-shelf AI tools.” Doing so can save “time and money by reducing dependence on outside translation services and specialized software.”
Human-in-the-loop frameworks have championed human translators’ use of AI as a tool, rather than as a replacement, but New Jersey’s training materials seem to encourage state agencies to take on translation as an in-house effort.
The rate of adoption by state agencies, and the scale of impact on language services providers, are both hard to predict, but they could be far-reaching. USDR reported that 15% of the US workforce (20m workers) have limited English proficiency (LEP).
