---
id: "workshop:beyond-translation-human-centered-language-access-with-ai"
type: "InnovateUS workshop"
title: "Beyond Translation: Human-Centered Language Access with AI"
url: "https://innovate-us.org/beyond-translation-human-centered-language-access-with-ai"
date: "2025-09-15"
duration: "60 minutes"
series: "Artificial Intelligence for the Public Sector"
instructors: ["Barbara Niveyro","Ruthie Nachmany"]
recording: "https://www.youtube.com/watch?v=E5UysrWWZ4E"
transcript: "automatic, from the recording"
source: "InnovateUS (innovate-us.org)"
licence: "CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org"
---

# Beyond Translation: Human-Centered Language Access with AI

InnovateUS workshop · 15 September 2025 · 60 minutes

How can public agencies ensure that AI-powered translations meet the real needs of the communities they serve? This workshop explores how a human-centered approach—grounded in plain language, bilingual co-creation, and machine translation supervision—can transform language access in government services.

Participants will learn from a real-world case study: how the New Jersey Office of Innovation collaborated with Spanish-speaking residents and frontline staff to redesign the Unemployment Insurance intake process, resulting in dramatic improvements in access, completion rates, and public trust.

This session will offer public sector professionals  tools and frameworks for incorporating AI and machine translation (MT) into their digital services—without losing sight of cultural nuance, accessibility, or equity.

This session is ideal for public servants in content design, digital services, equity and inclusion, user research, translation/localization, and IT. No technical background in AI is required.

## Learning goals

- Explain the difference between AI and machine translation and identify when and how to use each in service of language access
- Apply a 6-step framework for human-centered machine translation that includes plain language writing, bilingual review, and effective use of open MT tools
- Evaluate MT tools using key criteria such as integration capabilities, glossary support, and translation quality indexes
- Identify common technical and linguistic challenges in digital translation workflows and develop strategies to address them
- Promote inclusive and participatory practices by involving bilingual community members and frontline staff throughout the content creation process

## Speakers

- **Barbara Niveyro**, Bilingual Content Design Lead, Office of Innovation for the State of New Jersey
- **Ruthie Nachmany**, Product Manager, Office of Innovation for the State of New Jersey

Part of the series [Artificial Intelligence for the Public Sector](https://innovate-us.org/workshop-series/artificial-intelligence-for-the-public-sector).

Recording: https://www.youtube.com/watch?v=E5UysrWWZ4E

## Transcript

Automatic transcript of the recording: names and terms can be misheard, and the times can be a few seconds off. Each time links to that moment of the video. The transcript alone: https://innovate-us.org/beyond-translation-human-centered-language-access-with-ai.txt (plain text) and https://innovate-us.org/beyond-translation-human-centered-language-access-with-ai.json (every segment with its start and end second).

**Ruthie Nachmany**

[0:03](https://youtu.be/E5UysrWWZ4E?t=3) All right, let's get started. Can you all hear me okay? Okay, great. I see Barbara gave me a thumbs up, so that's good. Thanks, everyone. Wow, a lot of thumbs up. Thank you. Thank you so much to everyone for coming, and I just wanted to say thank you to everybody here for all the work you're doing on language access and improving government services with technology, AI, and human-centered practices. I just want to recognize – I know we have people who have been doing this work for many years, and I did see that someone in the chat – it's their first day of doing this work, so shout out to Victoria from D.C. And thank you so much to Innovate U.S. We appreciate your work so much at the New Jersey Office of Innovation. We learn a ton from our peers through the workshops that you lead, so I just wanted to say thank you to the whole team at Innovate U.S. Today, we're going to be talking about amplifying human-centered language access through machine translation and AI. Again, I'm Ruthie. I'm here with my colleague, Barbara. What I really want to emphasize kicking off this talk is that everything that we do at the New Jersey Office of Innovation, especially with AI, starts with a human-centered collaborative approach. For us, this collaboration is with New Jersey residents, bilingual support staff like call center agents, community-based organizations who work with bilingual and multilingual residents, and all of those partners collaborate with our team and teach us about how we can be doing a better job of making government services easier for everybody to understand and take opportunity of. And we use AI to scale and expand this fundamental work, but it always starts with that foundation. This is how we're going to be talking about today's workshop. We're going to go through a case study around unemployment insurance, our strategy and framework for language access beginning in Spanish, some best practices, choosing a tool, and how you can get started today. So again, we're going to start by talking about a human-centered approach to modernizing unemployment insurance in the state of New Jersey. And specifically, we're going to focus, we're going to talk a little about what we did in general around unemployment insurance and

[2:08](https://youtu.be/E5UysrWWZ4E?t=128) then specifically around how we improved the experience for Spanish speakers in New Jersey. So starting with the problem, you can see here's some quotes that we heard from people when we kicked off this project around the challenges Spanish speakers faced when trying to get access to unemployment insurance. I'll just read these out loud. I think it's helpful to hear. A call center agent let us know, I noticed that a lot of Spanish speakers are paying a third-party agency when they should be able to make a claim for free by themselves online. And then this is a translated quote. I don't know if you've seen what it's like to fill out a form in the USA. The truth is I prefer that a train run me over. I can't read it. It's so much that they make you dizzy. So we don't want that. I'll talk a little about what we did to improve this experience. If you can go to the next slide, Barbara. Great. So and again, we did all of this without AI. And then we'll talk about how we use this experience so that we could bring these insights to the work that we do to scale human-centered translation with AI. So in April 2024, we launched a new unemployment insurance application. This had 20% fewer questions. It reduced application time by 84%.

[3:13](https://youtu.be/E5UysrWWZ4E?t=193) It used to take people about four hours on average to fill out the application. And that takes them about 25 minutes on average. We cut call center wait times. This is the one I really want to highlight for this presentation. We achieved 100% completion time parity for English and Spanish speakers. And in a moment, I'll talk about some of the ways we did that. We also decreased claims requiring manual review by 14%. And we boosted post-application submission follow-through by 20%. And you can imagine that all of these support one another. If it's easier to fill out an application, you're more likely to fill it out in a way that is accurate and requires less manual review and follow-up. People can get the benefit faster. And again, we had an opportunity to really transform the unemployment insurance process end-to-end in the state of New Jersey. And like everything else that we do, we go deep in collaboration with one program in the state. And then we take out learnings and bring that out to many programs. But yeah, let's zoom in a little bit how we achieved 100% completion time parity for English and Spanish speakers. So this was a very kind of deep effort from a lot of folks, including eight months of manual content co-creation. We collaborated through 15 native Spanish speakers who live in New Jersey. And as I mentioned, also with support center staff, with community-based organizations, a really diverse group of folks, nine Latin American countries were represented, also five bilingual creatives who collaborated with the people I mentioned and kind of really tried to bring out their insights and figure out how we could translate that into the new application that we were developing. And again, this is serving 900,000 Spanish-speaking workers in the state of New Jersey.

[4:54](https://youtu.be/E5UysrWWZ4E?t=294) And then shortly after we released the new unemployment insurance application, New Jersey passed the 2024 language access law. I'll just walk through this quickly. So this requires translation of an agency's vital documents into the top seven languages spoken by individuals with limited English proficiency in New Jersey. And the key idea is really meaningful access, not just auto-translation. So as more tools are available for us to use, more technologies where we can do translation easily, really centering that human lived experience and that what we're translating really conveys the essential meaning of what the original text intended. As with everything we do in translation, this also helps us improve the plain language English content. So there's like this great flywheel that gets created. Many states and municipalities have passed similar laws or actively discussing them. So some of you may also be working through language access laws. And even if you aren't, I think this is kind of the fundamental nugget of this work is important for us all to think about. And based on the work we did with programs like unemployment insurance and with this language access law and the insights around the passing of this law, we've started to develop a framework around language access and what we've learned and scaling those insights out across the state. So I'm going to hand that over to Barbara now to talk through that a little more.

**Barbara Niveyro**

[6:13](https://youtu.be/E5UysrWWZ4E?t=373) Thank you, Ruthie. Hello, everyone. And just friendly reminder, as Ruthie said, we're going to share a lot of information, but we will have 30 minutes after. So if you want to start sharing questions in the chat, and then once we finish with a presentation, we can go back and read them. So a framework for language access is starting with Spanish. After this unemployment insurance project that Ruthie mentioned, these are our pillars co-create with bilingual frontline staff. For instance, for this project, collaborating with the people at the call center, they had so many insights about bilingual people calling, especially in this case, Spanish speakers. It helped us identify a lot of people who are bilingual and who are calling it help us identify a lot of issues with the vocabulary. They were very specific and very all the knowledge that these professionals share with us was very, it prepared us, even before we spoke with the residents, we already had a lot of information that was very valuable. On the other hand, embrace dialects. This is not going to happen with Spanish only.

[7:31](https://youtu.be/E5UysrWWZ4E?t=451) As you know, in New Jersey, we have Spanish speakers from different Latin American countries. And some words might be different, depending on the region. I always share the pineapple example, that if you're from South America, you're gonna say anana. And if you are from Central America, you might say pina. And if these pineapple example is essential for the process, for instance, a check, it's one of the words that it can be very different, depending on the dialect, it can be confusing. And if you actually take the time to address the difference, it was gonna create a lot of trust between your readers. And of course, you need to represent this with your glossaries and check the different options through UX research. And then let bilingual speakers lead and transcreate. Of course, we always want to keep the original version, the meaning of the original version. But the idea of transcreation is that, as a translator, we can have the freedom to slightly change what we are saying, if this is needed to ensure that representation is captured and to ensure that cultural nuances are being translated. Not all the cultures will need the same translation. So that's why transcreation gives you that freedom. This is a summary to tell us a little bit about the Office of Innovation. When we started with this project, we were very focused on research because it was a new topic for us, co-creation, and also basically having the big picture of what was the issue and what were our limitations. Right now, we are more focused on, and also that's why we were not using machine translation for our first project. Right now, that we know more about

[9:37](https://youtu.be/E5UysrWWZ4E?t=577) the technologies and what are the best ways we can use it to scale our translations, we are focused on technical resources and also socializing our learnings like this webinar or playbooks and guidelines. Without using AI, we translated all these projects except job protection, which was the first project that we translated with a translation management system. But manually or in a hybrid way, we translated for unemployment insurance, disability info hub, eviction help, family leave, direct file and file your state taxes, and an internalization list that I will be happy to share, which is basically a list with very common expressions and calls to actions and how we want to say them in English and how we want to say them in Spanish based on testing and also based on other services that have been translated for a while. Right now, we will be translating file your taxes, and that's going to be with machine translation.

[10:48](https://youtu.be/E5UysrWWZ4E?t=648) Not right now in a couple months, because right now we're more focused on the English version, but that's going to be with machine translation. So, here I'm going to share some best practices for using machine translation, and these are best practices for using machine translation with an open translation tool. I don't want to get too technical into a whole translation management system because that's more complex and not everybody has access to that. But as a starting point, just because AI and machine translation can be confusing for some people, I like to say that AI is your toolbox where you have so many different tools, and then AI is a specific screwdriver inside that toolbox. So, AI is the whole options you can have, and then machine translation is the automatic translation that AI helps you create.

[11:50](https://youtu.be/E5UysrWWZ4E?t=710) And then based on some tests that we made during our original project, we can always see that the best translation is the machine translation that is reviewed by a human. So, that's why, or for many reasons, we know that the human in the loop is so essential, even if you're going to scale your results with AI. But at the same time, we know the benefits of AI, and basically, because AI can help you make translation memories and accumulate data of your style, that's going to make the tool automatically write with the words that you are choosing. This example of do I want to say piña or anana, or like the most common example is that even for the Spanish version, we want to say New Jersey in English because in a localized translation, we realize that that's what people want to read, even for the Spanish version. So, we can train the tool to always say New Jersey in English, even if we are translating.

[12:59](https://youtu.be/E5UysrWWZ4E?t=779) At least for Spanish, right? And then, of course, we can move faster, and as you know, we have so many things that we need to translate, and Spanish is only the beginning. As Ruthie mentioned, we need to do this in seven languages. So, now I'm going to walk you through some steps that you could follow, that we recommend to follow when translating using open translation tools. So, the first step is write the text in English, in plain English, and test it. So, that's a big part, almost as big as the translation process is, if not bigger, actually. I tend to write in English first, and then I make the translations. In many projects, I have the privilege to write both versions, and writing the English version takes longer, and when I say that you need to ensure that the English version is ready, it's not that it only has to be in plain English. It also has to be tested. It has to be tested by your team, and also validated by your stakeholders. Of course, we always know there's a 20% of the content that is going to change, but if you have an 80% that has been reviewed and iterated, okay, that's when you can say that you're ready to start translating, and for those that don't know the federal translation guidelines, please find them. They are super simple, and they will help you to get started with plain English. This is a very simple step, but I don't want to take it for granted that you already know this. Be tidy, organize your folders, having a translation folder inside the team, and organize them by month is what makes sense to me, because you might have many versions of the same translation, so that's why I prefer organizing them by month. Some people might prefer to organize them by month and then by team.

[15:01](https://youtu.be/E5UysrWWZ4E?t=901) for your team, but remember that this is an important step. And then you can take a screenshot of this. This is a template where you can start copy-pasting the translations. If you use a translation management system in the past, also known as TMS, you will recognize this is a similar layout. So basically, I'm going to describe what we have here at the top. Basically, we want to incorporate basic information about the project. Who is requesting the translation? Is the English version written in plain English? Do we have a proofreader in Spanish? Is it ready? And do we have someone from our partner that will review the Spanish version? And then in the columns, the different content, the different type of text you will have in a layout, like checkboxes or h1, dh2. And then the Spanish column is where you're going to paste your machine translation.

[16:02](https://youtu.be/E5UysrWWZ4E?t=962) And then you're going to mark if this was reviewed by a human and if we have comments or things that we want to test later or if the state partner has notes. These are some basic prompts. I will show the same slide, but with an example. But we all know that when we are working with an AI assistant, we need to be specific about the instructions. We need to pretend it's someone that we are speaking with, that we are having a conversation with. And this is some prompts that I recommend for translation. Again, this might change if you are doing this for another language or for a very specific project. And also, the details will change on your style. But for us, we are saying that the language has to be Latin American Spanish. For instance, we don't have a lot of Spanish speakers from Spain, so that's why we want to make this clarification.

[17:04](https://youtu.be/E5UysrWWZ4E?t=1024) If your product is for people for a very specific dialect, you can be even more granular. The formality, we want to use a pronoun. Again, this might change depending on the product. For some products, we use the to pronoun because we want to be consistent with pre-existing parts of the same product. We care about being gender neutral, which is possible to do in Spanish, but it can be tricky. And then the format, we want the machine translation, the AI assistant to present the results into columns so that we can see the English version. So now I'm going to show a little demo. And then a follow-up question that you will see in the demo. But basically, a follow-up question that I tend to ask is, can you tell me the same but with a different dialect so that I can identify what terms might need a synonym. So I'm going to play the demo now. And I will voiceover it.

[18:09](https://youtu.be/E5UysrWWZ4E?t=1089) Copy-paste the prompt. This is the NJAI assistant. You can use the tool that you trust the most. I don't want to mention vendors, but you know them. You add the copy that you want to translate. You get the result. And then you ask the follow-up question. Based on this result, do you recommend synonyms I should consider to recognize the dialects? So it's giving me some alternatives. So now I transfer the Spanish version to my template. I made a mistake here. It's not English-English. It's English-Spanish, of course. And now I copy the alternatives because I want to analyze them. So I copy these alternatives. And for example, it's telling me I could consider saying seguro médico instead of cobertura de salud. And based on my experience, I'm like, you know what? I like seguro médico more than cobertura de salud.

[19:12](https://youtu.be/E5UysrWWZ4E?t=1152) So I'm going to replace it. I will use seguro médico. I think it's more aligned to healthcare coverage. And it might be more evergreen for Spanish speakers. And then the same thing. So now I'm like, okay, I'm done. I included that suggestion. And then now I'm going to say, oh, for multa, which means penalty, they are telling me we could also say penalidad. And since I don't know if multa or penalidad will be better, I will keep multa and penalidad. And then I will leave a reminder to see if we should say multa or penalidad or multa and penalidad because maybe it's a 50-50, but we want to test it. And then I'm marking that I'm the human made the review. So I'm going to move to the next slide, if I can stop the video.

[20:15](https://youtu.be/E5UysrWWZ4E?t=1215) Okay. And the final step is deliver the final translation to the appropriate team. And I have a question here for you if you want to share in the chat. How is your experience delivering translations or delivering content? Because that sometimes is very challenging. You have more than one version. You are working on the translation. And when you're working on the translation, the English version changed. So you need to incorporate that. So sometimes it can be a struggle if you want to share how you do it or if you have questions about how you do it. And then, of course, celebrate because you're handing off the translation. And based on that, just like five takeaways, start, continue, and finish with a human in the loop. I know this is hard, but if you can work with the same translator, at least for the entire project, that would be great. Because as you can see, we're making a lot of tiny decisions. That is really hard to have a new translator and, like, guess what were all these tiny changes.

[21:19](https://youtu.be/E5UysrWWZ4E?t=1279) And as you want to be consistent, of course, the AI tool can help you remember this. But if you can continue with the same person, that's going to save you a lot of time. Define your style and train your teams and tools. For example, in the example that I shared, we want to be gender neutral. And we also want to write in usted pronoun. In Spanish, you can talk in usted, tú, or vos. And maybe for your product, you want to use vos, which would be weird, but you might want to use vos. So define what's your style, because that's not only going to help your team, it's also going to train your tool. And then use more than one tool. If you are not working with a TMS, the complete translation environment, and you're using an open source translation, as you know, like, these technologies are very new, and they can hallucinate.

[22:21](https://youtu.be/E5UysrWWZ4E?t=1341) And also because some tools might be better with some tasks. Like, I like one tool because it's very good with dialects, but I prefer another tool because it's better with grammar. So I can back and forth. And then prepare for the translation domino effect. I don't mean to overwhelm you, but usually someone comes to me and they are like, hey, can you translate this page? And I'm like, yeah, but did you notice this page has 10 links to 10 other pages? What are we going to do about that? And then hand off a validation process. Like, you need somehow to validate that the translation makes sense. So this should be, we work with partners a lot. So we need someone to say like, yes, Barbara is not writing the lyrics of the Macarena song here. She's actually writing a serious translation that is loyal to the original English version. And I will pass the mic to Ruthie again to wrap up the presentation.

[23:25](https://youtu.be/E5UysrWWZ4E?t=1405) Thank you.

**Ruthie Nachmany**

[23:25](https://youtu.be/E5UysrWWZ4E?t=1405) Thanks, Barbara. And just want to say thank you so much to everyone who's been putting questions in the chat. Please feel free to keep putting in questions. And I see we have 158 people here. So I know there's a lot of cross-pollination and knowledge sharing. I've been answering some thoughts, and we can have a discussion, and I would love to hear from other folks too on how they're approaching things. So please feel free to comment in the chat as well. But anyway, let's wrap up with how to get started. If you go to the next slide, Barbara. So, yeah, the best way to learn is to learn by doing as quickly as possible. And I saw some conversation about who is a good team to lead this kind of work. We can discuss that further. Yeah, always helpful to understand at least initially, like, who the point person is going to be here and who's going to be supporting and informed. I want to highlight here kind of identifying within your context, what are the main languages and dialects spoken. And for a specific service, it may be different than for your entire state or municipality or kind of broader audience base.

[24:29](https://youtu.be/E5UysrWWZ4E?t=1469) It's something to understand a little bit more about as well, and obviously figure out who you can connect with in your team, support staff, community partners, et cetera, to get a sense of that. And if you go to the next slide, Barbara. Yeah, so in terms of pilot projects, so as Barbara talked about, sometimes there's downstream effects link, you know, always helpful to identify if there's, like, a small scoped initial project to work on just so you can kind of get a sense of the workflows and some of the places where, like, there's a bit of friction and then refine that with a smaller scope type of project. Recommend making a style guide or glossary that you can then upload to an AI tool used to train a machine learning model. If you remember that chart that Barbara showed, well, we saw that a human translation is always most effective at translating things in the way that we'd like to. A trained glossary of terms that you can upload to an AI tool is almost as good from what we've seen in our feedback from folks on those translations.

[25:31](https://youtu.be/E5UysrWWZ4E?t=1531) So, again, we really recommend that. And we will share a link to a glossary we've developed with unemployment insurance terms that you can use either for that kind of program or for similar, for programs that you are involved with that have similar terms. Again, once you have that, open up an AI system, upload the guide. Ask it to translate text. Get a translator to review those translations. Make sure that you incorporate any tweaks back into the model and then validate with user research, bilingual users, staff, and so on. Again, kind of bringing in those insights from various folks. And then we're going to share some resources. So, as I mentioned, we have a translation guide for unemployment insurance, but many benefits programs use similar terms. So, take a look at that. And we'll be sharing, I think, all the slides will be shared in the recording too. So, if you don't get a photo of this, you can look at the recording later. Also, when you translate, the layout of a page and the way you want to design content and display it may evolve. So, we have a playbook here around content design and language access and how the two can inform one another and glossary of internationalization terms you can use to kickstart your work as well.

[26:38](https://youtu.be/E5UysrWWZ4E?t=1598) And if you want to read a little bit more about some of the work we've done, you can check out these blog posts. And I think with that, we can open up the room for conversation.

* * *

Original page: https://innovate-us.org/beyond-translation-human-centered-language-access-with-ai · Licence: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org