Warning: Invalid argument supplied for foreach() in /home/customer/www/fahamuusaimmigration.com/public_html/wp-content/plugins/affiliate-ads-builder-for-clickbank-products/vertical_horizontal_carousel.inc.php on line 70
In 2019, Carlos fled Brazil with his sister and two nephews after his son was murdered in front of him by a local gang. Upon arriving in the US, he was separated from his family and detained in a US Immigration and Customs Enforcement (Ice) detention center.
Carlos, who is Afro-Indigenous, speaks Portuguese but does not read or write it. Staff at the Calexico, California, detention center spoke only English or Spanish. The staff used an artificial intelligence-powered voice-translation tool to interpret what Carlos was saying, but the system didn’t pick up or understand his regional accent or dialect. So Carlos spent six months in Ice detention unable to meaningfully communicate with anyone.
In that time, he had no clear idea of why he was being detained or where his family was. When he sought medical care for his high blood pressure and for Covid, the nurses had trouble understanding him, he said. Spanish-speaking fellow detainees helped to fill out his asylum application, but the translation tool they used failed to produce an accurate account. It didn’t recognize Belo Horizonte as the name of one of the cities Carlos had lived in, instead translating it literally to “beautiful horizon”. And in response to a question about the mistreatment he suffered, the application read: “YES THE GANGUE DO BURACAO TO SHOOT DEAD MY SON, IN THE POLICE I WAS SLAPPED.”
Carlos, for whom the Guardian is using a pseudonym so as not to compromise his residency application, said the language barrier was among the hardest challenges in his quest to seek refuge in the US. “I never imagined that would be the worst thing,” he said through a translator.
The US immigration system has said it will provide migrants with a human interpreter as needed. In reality, refugee organizations say many are frequently left without access to one. Instead, the various agencies that make up the US immigration system and even some refugee aid organizations increasingly rely on AI-powered translation tools like Google Translate and Microsoft Translator to bridge that gap.
The Department of Homeland Security (DHS) has set up several contracts with machine translation firms, including Lionbridge and TransPerfect Translations International Inc. Immigration officials at Ice have been instructed to use Google Translate to vet refugee applications. Customs and Border Protection, which is in charge of border management, has even developed its own app, CBP Translate, to help communicate with migrants.
But the US immigration system has come to rely on these translation tools without fully grasping their limits, said Ariel Koren, the founder of Respond Crisis Translation, a network of 2,500 people who provide urgent interpretation services.
“AI translation tools should never be used in a way that is unsupervised. They should never be used to replace translators and interpreters and they should not be used in high-stakes situations – not in any language and especially not for languages that are marginalized,” Koren said.
AI-powered translation tools are particularly unreliable for languages that are considerably different from English or are less comprehensively documented, said Damian Harris-Hernandez, the executive director of the Refugee Translation Project, another group that helps refugees with translations.
“It’s very tempting for a lot of organizations or companies to use machine translations,” Harris-Hernandez said. “But these discrepancies can void a whole [immigration] case.”
Problems with the translation tools occur throughout the asylum process, from border stations to detention centers to immigration courts, said several volunteers at Respond Crisis Translation. The CBP One app, which the Biden administration has mandated anyone seeking asylum to use to schedule an appointment with CBP before entering the country, is translated into only a handful of languages. And even in those translations, errors appear. The version of the FAQ section of the app in Haitian Creole, for instance, largely shows a string of letters with no spaces or the necessary accent marks.
Respond Crisis Translation volunteers say they have seen cases of asylum applications being denied because the translation tool interpreted an “I” in a refugee’s statement as “we”, making it seem as if it was an application for more than one person. They also recalled the case of a woman seeking asylum due to domestic abuse who described her abuser as “mi jefe” in her application. The woman was using the term colloquially to describe her father, but the translation service translated it literally to “my boss”. Her asylum application was denied.
“Not only do the asylum applications have to be translated, but the government will frequently weaponize small language technicalities to justify deporting someone,” said Koren, who used to work at Google Translate. “The application needs to be absolutely perfect.”
The volunteers say the translation challenges have been particularly troubling in the case of Afghan refugees. Major machine translation systems like Google Translate, for example, do not offer translations in Dari, one of two official languages of Afghanistan.
“Afghan languages are not highly resourced in terms of technology, in particular local dialects,” said Uma Mirkhail, the group’s lead for the 40-person team representing Afghan languages. “It’s almost impossible for a machine to convey the same message that a professional interpreter with awareness about the country of origin can do, including cultural context.”
Google Translate is available in 133 languages and the system is “rigorously” trained “to ensure every language we support meets a high standard”, said a Google spokesperson, Charity Mhende, but without sufficient data it can be difficult to develop language translation and speech recognition at scale, she said.
DHS, which encompasses CBP and Ice, did not respond to repeated requests for comment.
The failures of machine translations the volunteers encountered and the inability of AI to translate the world’s languages is a systemic problem, argue Paula Helm, an assistant professor in data science and ethics at the University of Amsterdam, and Gábor Bella, an assistant professor in the University of Trento’s department of information engineering and computer science.
AI systems are dependent on the data they are fed. The quality of that training data is determined by both its accuracy and its comprehensiveness.
Much like other AI tools, machine translation services reflect and perpetuate existing biases in society and global power and economic imbalances, the researchers said. Due to its colonial and imperial history, the English language, for example, is among the most recorded in the world. Thus, there’s no shortage of English-language data to be fed into AI systems. But when it comes to a language like Swahili, which is spoken by more than 80 million people across Africa, digital sources are much scarcer, according to the researchers. In fact, there are about as many Wikipedia pages in Swahili as there are in Breton, a language spoken by a little more than 200,000 people in a small region in France.
In addition, there are cultural nuances that can’t always be communicated in English, they said. “When you go through the filter of the English language and what can and cannot be expressed in the English language” there can be knowledge loss between languages, said Bella.
The researchers pointed to the example of the word “rice”. Many languages don’t have a generic word for rice and instead have words that specify whether it is raw, cooked or brown. So when translating “the rice is tasty” to Swahili through Google Translate, the result – “mchele huu ni kitamu” – means “this uncooked rice is tasty”.
Language is more than a series of words and their meanings; it’s a means to express cultural identity, and it’s how many communities make sense of the world, the researchers said. Without cultural context, machine translation systems will continue to prioritize a western worldview, making it nearly impossible to properly interpret the nuances of most non-English languages.
“These large language models are being developed by computer science communities, mostly funded by rich western universities where English is the default language and everything else is a niche topic,” Helm said.
While Bella believes that one potential solution is to ensure those contributing to the training data for various languages are native speakers who are given control over the data, some of the translators at Respond Crisis Translation are not convinced it will help, especially in crisis situations.
“Data is still data and a human is a human,” said Yaseen, a translator on the Afghan language team. “We are dealing with people who are traumatized and our approach is trauma-informed. As an interpreter, you cannot under-do or overdo [the translation], but at the same time, you should have empathy to convey their emotions and feelings and that is only possible with a human being.”
Six months after he was first detained, Carlos was finally connected with Respond Crisis Translation through a legal aid clinic.
Meeting his interpreter, Samara, was a relief, he said. He finally found someone who could understand and help him.
With her help, Carlos ultimately received asylum and is now applying for his residency.
“Having someone to speak my language and for the first time after so many months was the beginning of hope,” he said.