Volunteer network of interpreters hopes to make refugees' languages more accessible via AI
NEW YORK — They may be Tigrinya speakers fleeing the authoritarian Eritrean government's indefinite military service policy. Or Rohingya people escaping ethnic violence in Myanmar. But refugees navigating resettlement often face a shared hurdle: poor machine translations and a short supply of interpreters knowledgeable in their less-serviced languages.
Tarjimly, a Google-backed nonprofit described as "Uber for translators," aims to help asylum seekers clear that hurdle. Through a new artificial intelligence partnership, Tarjimly trains outside large language models while allowing its volunteers to respond more urgently to needs for translators. It's a feedback loop where humans teach the nuances of each language to the machines by sharing data from one-on-one calls and correcting automated translations.
And it's this uniquely human realm of language that Tarjimly co-founder Atif Javed believes exemplifies the ever-tricky balance between individuals' ingenuity and technological advancement. He says it's the needed personal touch that shows why AI's rapid development shouldn't generally stoke widespread fears.
Languages popular in the Global South — such as the Dari and Pashto commonly spoken in Afghanistan, home to one of the world's largest protracted refugee crises — have the worst quality coverage, according to Javed. He feels well positioned to supplement the internet's English-dominated information troves that train services like Google Translate with his mobile app's more diverse data sets.
Tarjimly connects refugees with on-demand interpreters, who can communicate during meetings with social workers, immigration officials and doctors, and records the encounters for AI training. To comply with patient privacy protections, Tarjimly anonymizes the conversations on its app. Javed said the nonprofit also has on option for "no record" sessions where none of the data is stored for alternative uses.
Many of its 60,000 volunteers are multilingual refugees themselves who more intimately understand not only their counterpart's native tongue but also the crisis that brought them there, according to Javed.
Among them is Roza Tesfazion, a 26-year-old Eritrean refugee who works professionally as an interpreter for the United Kingdom's government. Fluent in Amharic and Tigrinya, she studied English and Swahili to help her immigrant family overcome language barriers when they first moved to Kenya.
Tesfazion said she translates at no cost because she knows "how emotional it is" for the people on the other side of her sessions.
"You have to have that touch of human emotions to it," she said.
Tarjimly's founders say their mission's sensitive nature lends itself to nonprofit status more than a corporate structure. Users arrive in very vulnerable positions, and the nonprofit works with established humanitarian groups including Catholic Charities, the International Rescue Committee and the United Nations' International Organization for Migration.
The work requires a level of trust that would have been difficult to earn in a "for-profit, competitive world," according to Javed. "The underlying engine of our success is the community we've built."
That community, however, also has room for artificial intelligence. A $1.3 million grant from Google.org has enabled a "First Pass" tool that gives an instantly generated translation for human volunteers to revise. A new information hub will open up its language data for partners, including Google, in early 2025.
But refining a more diverse library of languages will require conversational data at a scale much broader than Tarjimly can likely provide on its own, according to Data & Society researcher Ranjit Singh.
Singh, who studies the social implications of automation and inclusive digital solutions, said translation services will always need a "real person in the middle."
"There is one part of it which is translation and another part of it which is just trying to understand somebody's life situation," he said. "Technologies help us do some of this work. But at the same time, it's also fairly social."
Tarjimly was inspired by Javed's time volunteering with Arabic speakers at refugee camps in Greece and Turkey after graduating from the Massachusetts Institute of Technology and working in Silicon Valley. A Muslim American whose family immigrated to the United States in 2001, Javed said he was reminded of his own childhood translating for his refugee grandmother.
His lived experience is one reason why Elevate Prize Foundation CEO Carolina Garcìa Jayaram said her organization awarded $300,000 last year to Tarjimly. That "proximate leadership" helps nonprofits better understand developments like artificial intelligence that "can be both cause for excitement and trepidation," Jayaram said. The risk-averse philanthropic sector may be slow to catch up with disruptive new technologies, she noted, but shouldn't ignore their positive applications.
"It's a great example of how not to get stuck in that bogeyman complex about AI," she said. "To go to leaders who are closest to those issues and say, 'How would AI unlock the possibilities and opportunities for your organization?'"
Tarjimly, a Google-backed nonprofit described as "Uber for translators," aims to help asylum seekers clear that hurdle. Through a new artificial intelligence partnership, Tarjimly trains outside large language models while allowing its volunteers to respond more urgently to needs for translators. It's a feedback loop where humans teach the nuances of each language to the machines by sharing data from one-on-one calls and correcting automated translations.
And it's this uniquely human realm of language that Tarjimly co-founder Atif Javed believes exemplifies the ever-tricky balance between individuals' ingenuity and technological advancement. He says it's the needed personal touch that shows why AI's rapid development shouldn't generally stoke widespread fears.
Languages popular in the Global South — such as the Dari and Pashto commonly spoken in Afghanistan, home to one of the world's largest protracted refugee crises — have the worst quality coverage, according to Javed. He feels well positioned to supplement the internet's English-dominated information troves that train services like Google Translate with his mobile app's more diverse data sets.
Tarjimly connects refugees with on-demand interpreters, who can communicate during meetings with social workers, immigration officials and doctors, and records the encounters for AI training. To comply with patient privacy protections, Tarjimly anonymizes the conversations on its app. Javed said the nonprofit also has on option for "no record" sessions where none of the data is stored for alternative uses.
Many of its 60,000 volunteers are multilingual refugees themselves who more intimately understand not only their counterpart's native tongue but also the crisis that brought them there, according to Javed.
Among them is Roza Tesfazion, a 26-year-old Eritrean refugee who works professionally as an interpreter for the United Kingdom's government. Fluent in Amharic and Tigrinya, she studied English and Swahili to help her immigrant family overcome language barriers when they first moved to Kenya.
Tesfazion said she translates at no cost because she knows "how emotional it is" for the people on the other side of her sessions.
"You have to have that touch of human emotions to it," she said.
Tarjimly's founders say their mission's sensitive nature lends itself to nonprofit status more than a corporate structure. Users arrive in very vulnerable positions, and the nonprofit works with established humanitarian groups including Catholic Charities, the International Rescue Committee and the United Nations' International Organization for Migration.
The work requires a level of trust that would have been difficult to earn in a "for-profit, competitive world," according to Javed. "The underlying engine of our success is the community we've built."
That community, however, also has room for artificial intelligence. A $1.3 million grant from Google.org has enabled a "First Pass" tool that gives an instantly generated translation for human volunteers to revise. A new information hub will open up its language data for partners, including Google, in early 2025.
But refining a more diverse library of languages will require conversational data at a scale much broader than Tarjimly can likely provide on its own, according to Data & Society researcher Ranjit Singh.
Singh, who studies the social implications of automation and inclusive digital solutions, said translation services will always need a "real person in the middle."
"There is one part of it which is translation and another part of it which is just trying to understand somebody's life situation," he said. "Technologies help us do some of this work. But at the same time, it's also fairly social."
Tarjimly was inspired by Javed's time volunteering with Arabic speakers at refugee camps in Greece and Turkey after graduating from the Massachusetts Institute of Technology and working in Silicon Valley. A Muslim American whose family immigrated to the United States in 2001, Javed said he was reminded of his own childhood translating for his refugee grandmother.
His lived experience is one reason why Elevate Prize Foundation CEO Carolina Garcìa Jayaram said her organization awarded $300,000 last year to Tarjimly. That "proximate leadership" helps nonprofits better understand developments like artificial intelligence that "can be both cause for excitement and trepidation," Jayaram said. The risk-averse philanthropic sector may be slow to catch up with disruptive new technologies, she noted, but shouldn't ignore their positive applications.
"It's a great example of how not to get stuck in that bogeyman complex about AI," she said. "To go to leaders who are closest to those issues and say, 'How would AI unlock the possibilities and opportunities for your organization?'"