How language translation technology is jeopardizing Afghan asylum-seekers

Read the Full Transcript

• John Yang:

  Since U. S. forces left Afghanistan in 2021, the humanitarian situation there has gotten worse by the day. This past week, the Biden administration took steps to allow some Afghans who fled after the Taliban takeover to extend their temporary stays in the United States. But for those still applying for asylum, it's proving increasingly difficult. Ali Rogin has more on how some claims are getting lost in translation.

• Ali Rogin:

  In Afghanistan today, over 28 million people, two-thirds of the population, require humanitarian assistance. Six million people are living in near famine conditions. Women and girls remain incredibly vulnerable under Taliban rule, all of these factors leading many Afghans to seek asylum here in the US.

  Now, machine learning technology is being used to translate the dozens of languages spoken in Afghanistan. But a new investigation by the news organization Rest of World details how various forms of language translation technology are creating computer errors that put asylum seekers lives in danger.

  We recently spoke with a translator who works with Afghan asylum seekers. We've concealed her identity for her safety and the well-being of her family.

• Woman:

  The raise in machine translation and apps are not only costing translators their jobs, but quite literally jeopardizing asylum cases. Like a recent example of this is a Dari translation done by a machine mistranslated actually, I as we. This created an inconsistency between the asylum seeker initial interview and what was written in their asylum application. This consequently was enough for the judge to reject the case.

• Ali Rogin:

  Joining me to discuss how translation technology is putting asylum seekers at risk is Andrew Deck, who wrote the investigative story, and Leila Lorenzo, policy director at Respond Crisis Translation, which provides translation services for migrants and refugees. Thank you both so much for joining me.

  Andrew, I'm going to start with you. Why does the U.S. government say it's using this technology, and what do we know about how widespread its uses.

  Andrew Deck, Reporter, Rest of World: You know, in the U.S. as of December 2022, there was a backlog of 1.6 million asylum applications. And, you know, one way that this technology is framed is a way to speed up the processing of these applications.

  But what our reporting bore out is that it's also a way of kind of cutting corners in terms of cost, especially, that has long tail impact on the quality of translations, and that becomes concerning quickly when we're dealing with the safety and security of incredibly vulnerable communities like Afghan refugees.

• Ali Rogin:

  And Leila, you are indeed part of this translation community. So what are you seeing? What are some of the problems that the people you're working with are experiencing in using this AI technology?

• Leila Lorenzo, Policy Director, Respond Crisis Translation:

  I think that we've seen a variety of problems with the translation technology. For example, with the CBP One app that's used at the border, there's only a few number of languages that can be used English, Spanish, and in some cases Haitian Creole. And we have received reports here at Respond Crisis Translation that the Haitian Creole programming, for example, is mistranslated because they only use machine translation.

  And we've seen several instances here at Respond Crisis Translation of mistranslations in critical asylum documents that can result in seriously jeopardizing the case or completely invalidating them.

• Ali Rogin:

  Andrew, what are some of the most common problems that happen when these programs are used to help with the Afghan asylum process?

• Andrew Deck:

  Machine translation really struggles with idioms, cultural nuances, different dialects and slang. And those problems are only exacerbated with Pashto and Dari.

  You know, one translator I spoke with at Respond Crisis Translation told me that machine translation tools he'd tested really struggled with things like military rank in Pashto and Dari. That's concerning because so many refugees entering the asylum review process in the U.S., you know, they worked closely with U.S. military and U.S. allied forces in Afghanistan, which is essential to their claim of a, say, credible fear of persecution.

  So these kinds of mistakes, they may seem trivial when we're talking about something like Google Translate, but we know that asylum review is a rigorous process where some of these small inconsistencies, they can properly jeopardize a claim.

• Ali Rogin:

  Leila, as it happens, many of the people that were trying to flee the Taliban in the summer of 2021 were people who were providing translation and interpretation services for the United States. Can you tell me about what your organization is doing to help those people?

  And also, is it a missed opportunity for the United States government to not try to take advantage of this brain power and at least help them help us in translating some of these Afghan languages?

• Leila Lorenzo:

  I would shy away from framing it as a missed opportunity. Rather, I think it's actually a failure on the part of the United States government. This is a missed opportunity to provide paid work to families in Afghanistan where translators may have multiple members, like up to 11 family members that they are supporting one income and that's what we've directly observed at Respond Crisis Translation.

  So our organization is working very hard to provide paid work, and not only paid work, but finding ways to circumvent what can only be described as a collapsed economic system. And if that's something that Respond Crisis Translation can do on a very limited budget, that is something that bigger companies and the United States government can figure out a way to do.

• Ali Rogin:

  And to both of you've clearly laid out some of the problems with using this technology. Are there any solutions here, or is the overarching response simply not to use these automated translation services?

• Andrew Deck:

  I mean, AI is largely an unregulated sector of the tech industry right now, and that includes machine translation. As a result, we're often relying and expecting right now for private companies to monitor harm and self-regulate the use of their own products.

  But ultimately, I think we need to put the onus on agencies and aid organizations to more closely investigate and interrogate the use of this technology and not necessarily private companies that may be competing over lucrative government contracts.

• Leila Lorenzo:

  Our position at Respond Crisis Translation is that we do not use machine translation, and especially in the case of Dari which is an underresourced language, you'll notice. On Google Translate, it is subsumed under Persian, and Persian is a language with a dialect continuum. There is the Persian that is spoken in Iran, there is the Persian that is spoken in Afghanistan, and therefore I find it completely inappropriate. And that would be the position that translators who are working in context crisis, such as Respond Crisis Translation, would hold.

• Ali Rogin:

  Andrew Deck with Rest of World and Leila Lorenzo with Respond Crisis Translation. Thank you both so much for joining us.

• Andrew Deck:

  Thank you, Ali.

• Leila Lorenzo:

  Thank you.