Estimated read time: 8-9 minutes
Editor's note: This is part of a KSL.com series looking at the rise of artificial intelligence technology tools such as ChatGPT, the opportunities and risks they pose and what impacts they could have on various aspects of our daily lives.
SALT LAKE CITY — Utah Driver License Division Director Christopher Caras was terrified when the state first floated the idea of translating the driver's license test into multiple languages.
Other states had already undertaken the endeavor and estimated the project would cost $50,000 per language, per iteration of the test. Caras had sticker shock. Meanwhile, a 2022 law directing the division to offer the test in five new languages by 2023 only allotted $60,000 total (and individuals do not pay more to take the test in a language other than English).
Artificial intelligence translation technology became the division's saving grace, allowing it to translate the test into five languages at the cost other states were paying for a single language.
The division uses a combination of two technologies from Amazon, except for a few languages that aren't available through the service. In those cases — or when a test question appears to be off based on data from test takers — the division will pay a human translator. So far, the division has translated the test into Spanish, Tongan, Vietnamese, Mandarin and Portuguese. It hopes to add another five languages each year.
"When you're talking languages, there's so many dialects and nuances that you're really never going to get everybody to agree, even if you ask all the certified translators to come in," Caras said. "Our aim is to get as close to that as we can by utilizing the technologies so that the cost is fiscally responsible to the state and the agency."
The Driver License Division is a leader among state agencies when it comes to using AI translation technology to better serve residents with little to no English skills. Other state and local government entities appear to only have options to translate webpages into another language or have the content spoken aloud for those who are visually impaired. The state is also looking into American Sign Language translation, which would take the form of a video of an ASL interpreter translating on screen.
Exploring such options may become necessary as the state becomes increasingly more diverse, as evidenced by the fact that over 120 languages are spoken in Utah.
"As a state, we are kind of at the strategic level, focusing on analyzing appropriate use cases and separating facts from fiction when it comes to the claims about the dependability of AI versus what really is safe and secure to use," said Rich Saunders, Utah's first chief innovation officer. "We are literally right now focused on a broad strategic level about how to use AI — and there's a particular interest in closing the language gap."
Saunders added that the lower cost of AI translation, as well as increased speed compared to human translation, are some of the biggest benefits of using artificial intelligence. Nubia Peña, the governor's senior advisor on equity and opportunity, said closing the language gap is an important part of the state's commitment to creating a climate that is inclusive and welcoming.
"In addition to contributing to Utah's economic success, our multicultural and linguistically diverse Utahns are enhancing the cultural vibrancy of our state," she said over email. "Integrating languages into service delivery models allows us to increase engagement opportunities with our constituents, helps improve services offered by creating greater accessibility so we can strive to better reach underserved communities."
How does AI translation technology work?
Adam Youngfield, a translation expert with over a decade of experience, said there's often a lack of understanding about how AI actually works.
"There's the general kind of folk understanding of what AI is, which is the sense that it's almost a human-like thing that can think and it can reason and it can come up with ideas and creations," he said. "That's not AI as it is right now."
Two of the most prominent types of AI technologies in the translation industry are neural machine translation, such as Google Translate, and large language models, like ChatGPT.
The first large-scale neural machine translation system was adopted in 2015 and it's had a big impact on translation ever since. These types of neural networks can produce — especially when it comes to more sophisticated technology than Google Translate — very accurate translations and have the ability to "learn" and improve over time. However, neural machine translations can also lack fluency.
In contrast, large language models are fluent but may not be very accurate. Although ChatGPT can understand and generate text in a variety of languages, it was primarily designed to understand and generate text in English.
There's the general kind of folk understanding of what AI is, which is the sense that it's almost a human-like thing that can think and it can reason and it can come up with ideas and creations. That's not AI as it is right now.
– Adam Youngfield, translation expert
The practical uses of neural machine translation range from on-demand translations of websites to live translation of text or speech (although Youngfield said speech-to-speech translations are still being developed).
The applications of large language models could be much broader. Not only could a user translate a webpage, but they could also ask follow-up questions or request additional resources in their preferred language.
"It's much more conversant," Youngfield said. "We're barely beginning to scratch the surface on what ChatGPT is capable of. ... As the technology continues to expand, you could envision the government even making their own version and deploying that from their website. From that point of view, it could even become like a virtual assistant for users of the website."
A number of Utah startups are exploring the intersection of AI and translations, said translation/localization professor and consultant Adam Wooten. He pointed to Addavox — which uses AI for automatic transcriptions, captioning and voice dubbing in over 100 languages — and OneMeta AI — which includes live translations in 82 languages.
Ethical concerns of governments using AI for translation
Both Wooten and Youngfield cautioned against using AI indiscriminately and without human checks. The two experts encouraged governments to think through what types of materials can be machine-translated and which require a human to either interpret directly or at least edit a machine translation.
"If you or I were to go to ChatGPT and rely on it for information, maybe in some cases that might be innocuous and maybe it's not like going to pose a real threat or danger to us. But if we were to ask it for information that might have an impact on our health or our well-being and it misled us in some way that caused us harm, then that poses an ethical challenge," Youngfield said. "For most publication purposes — especially where you want to maintain credibility and you want to have a high-quality product — you're not just going to publish raw machine translation, you're going to want to have someone review it."
Wooten pointed to a recent court case where police used a translation app to interrogate an individual who only spoke Russian. The defendant was acquitted. Wooten expects that a combination of human and machine translation will become more and more common, although he added a machine translation alone may be better than nothing at times.
"You definitely want to evaluate — whatever it is that we're translating — how much risk is there here? How much of a chance is there that if we mess something up, it's going to cause some hurt, pain, loss of liberty or loss of life?" Wooten said. "You want to be very wary of using that in these more risky situations. Sometimes risky situations are even created out of nothing, so it's often good to make sure that you have professionals advising you."
Caras, the Driver License Division director, recommended that other government entities engage the community as they explore translation options. In identifying which languages to prioritize for translation, for example, Caras said the division met with Pacific Islander leaders who expressed that the greatest need in the community was Tongan.
"Work with them when you get to those areas where you're kind of splitting hairs," he said. "Our experience was very positive. They were willing to step up and say, 'If anybody has to wait, we'll wait because we see a larger need here.' And that was immensely helpful."
Wooten added that politicians and government leaders also need to consider the limitations of AI when it comes to cultural fluency and how the sole use of AI translations might come across.
"Certain constituents, perhaps bilingual immigrants, might say, 'Oh, this politician, they don't really care about us. They're just using a machine. We're not good enough for them to hire one of us as a human translator or interpreter,'" Wooten said. "Politicians who wish to support their diverse constituencies might consider supporting actual human translation and interpreting services as this industry is, by nature, a very diverse industry employing a lot of bilingual immigrants."
