AI translation is jeopardizing Afghan asylum claims

In 2020, a Pashto-speaking Afghan refugee's asylum application was denied by a U.S. court due to discrepancies between her written statement and initial interviews. Crisis translator Uma Mirkhail discovered that an automated translation tool had incorrectly changed 'I' pronouns to 'we' in the woman's written statement, creating inconsistencies about whether she experienced a particular event alone or with others. This translation error was significant enough for a judge to reject the asylum claim. The incident highlights broader risks with machine translation tools being used in immigration systems, particularly for Afghan languages like Pashto and Dari, which are considered 'low-resource' languages with limited training data compared to dominant languages. U.S. Citizenship and Immigration Services officers have been instructed to use Google Translate for vetting asylum applicants' social media accounts, and major translation companies with federal immigration contracts advertise machine translation services. Translation organizations working with refugees report that machine translation tools struggle with cultural context, idioms, and specialized terminology crucial for asylum applications, with one expert estimating 40% of asylum decisions could be incorrect if relying on machine translation.