Government forms can often be a maze of unfamiliar terminology, and that might be doubly true if the paperwork isn’t in one’s native language.
That’s why the state of New Jersey has turned to generative AI to make its unemployment insurance application process easier for both English and Spanish speakers, with support from the nonprofit US Digital Response (USDR).
That team now wants to help other governments do the same with a new open-source toolkit designed for states developing their own custom translation assistants using off-the-shelf LLMs. Google’s philanthropic arm, Google.org, has also lent its backing.
New Jersey’s own generative AI translation tools, part of a broader modernization of the state’s unemployment system, are internal and not consumer-facing, meaning that a human employee always has oversight of the process.
The state’s previous unemployment insurance application system, which had been in place since the mid-1990s, offered only a Google Translate dropdown for non-English speakers, according to Gillian Gutierrez, senior advisor and director of NJ Unemployment Insurance Modernization.
But as anybody who’s used classic Google Translate knows, its output can be clunky and especially risky when dealing with legally precise language. One of the first steps in the overhaul process was to define common terms like “able and available” and the difference between “fired” and “laid off” in plain language, then translate to Spanish.
The team also pulled in the knowledge of call center agents who were on the front lines of dealing with these questions. “We poured in their knowledge of the different ways that they might have to explain those terms, given if the person is of Ecuadorian descent, or the person is of Dominican descent, or the person is of Mexican descent,” Gutierrez said.
Keep up with the innovative tech transforming business
Tech Brew keeps business leaders up-to-date on the latest innovations, automation advances, policy shifts, and more, so they can make informed decisions about tech.
Gutierrez said the team then worked with USDR to turn this glossary of terms into an LLM-based system that staffers could call upon to create more nuanced translations.
“We built the prompts of how you get the AI tool to give you back, based on the language inputs of the complicated terms, as well as the feel, the structure, the tone that you want the application to have as a government service, to give you back a translation that is as close to not needing a human review as possible,” Gutierrez said.
A blueprint for other agencies: Marcie Chin, language access product delivery manager at USDR, said the training material release this week is designed to help other states and municipalities replicate New Jersey’s language accessibility work. The nonprofit aims to provide tools that require minimal technical know-how and financial resources and are compatible with foundation models like Anthropic’s Claude, Google’s Gemini, or OpenAI’s ChatGPT.
“This novel approach de-risks translating complex government services for both governments and residents who historically struggle to access critical public services,” Chin said in an email. “Any agency can use these free, open-source resources to create custom AI translation assistants for critical services like unemployment benefits, delivering meaningful results without requiring technical expertise or heavy investment.”
Chin also said the tools are geared for the greater human oversight needed for government services than might be necessary in a business setting.
“Unlike the private sector, government cannot afford to fail fast—especially with an emerging technology like generative AI that is still evolving so quickly,” Chin said. “Investing thoughtfully and preserving human oversight is critical to using this powerful technology for improving the delivery of public services in a sustainable way.”