AI translation is jeopardizing Afghan asylum claims

A Pashto-speaking refugee's asylum claim was denied by a US agency for a discrepancy between oral and written recount of an event allegedly due to an error of their automated translation tool which swapped pronouns of her written statement from "I" to "we".

Risk classification

• Primary risk domain: 1 Discrimination & Toxicity
• Primary risk subdomain: 1.3 Unequal performance across groups

The translation tool performed poorly on Pashto, a low-resource language, compared to high-resource languages, leading to an unequal and highly detrimental outcome for an Afghan refugee.

Additional risk subdomains

• 7.3 Lack of capability or robustness: The automated translation tool failed to perform reliably, introducing a critical pronoun error that altered the meaning of the asylum application.
• 5.1 Overreliance and unsafe use: Immigration authorities and workflows relied on automated translation tools in high-stakes legal contexts without sufficient human oversight.

Causal factors

• Entity: AI
• Intent: Unintentional
• Timing: Post-deployment

The incident was caused by an unintentional translation error ('I' swapped to 'we') produced by a deployed machine translation tool during post-deployment use.

EU AI Act risk tier

• Risk tier: 2 High Risk

Risk Level 2: High Risk. The report describes the use of AI translation tools in 'high-risk government decision-making,' specifically migration and asylum applications, which has significant implications for fundamental rights.

AI system and alleged parties

• AI system: automated translation tool
• AI purpose: Translation; Writing Assistant
• Behaviour type: Tool
• Alleged developer: unknown
• Alleged deployer: US Citizenship and Immigration Services
• Alleged harmed parties: Pashto-speaking asylum seekers, Dari-speaking asylum seekers, anonymous Pashto-speaking refugee

Harm severity

Highest direct severity in any category: Substantial. Severity is scored from Negligible to Catastrophic in each harm category, for harm the reports describe as caused directly or indirectly by the AI system.

• Physical: direct Negligible, indirect Negligible
• Infrastructure: direct Negligible, indirect Negligible
• Property: direct Negligible, indirect Negligible
• Financial: direct Negligible, indirect Negligible
• Environmental: direct Negligible, indirect Negligible
• Malicious content: direct Negligible, indirect Negligible
• Differential treatment: direct Minor, indirect Negligible
• Civil rights: direct Minor, indirect Negligible
• Democracy: direct Negligible, indirect Negligible
• Privacy: direct Negligible, indirect Negligible
• Psychological: direct Negligible, indirect Negligible
• Epistemic: direct Minor, indirect Negligible
• Child sexual exploitation and abuse: direct Negligible, indirect Negligible

Differential treatment

Reported: The report explicitly describes differential treatment resulting from unequal performance of translation tools across different language groups.

Directly caused: The refugee faced unequal treatment and a rejected application because the translation tool performed poorly on Pashto compared to high-resource languages.

Indirectly caused: N/A

Inferred additional harm: Other refugees speaking low-resource languages likely face systematically higher rates of application rejection and scrutiny due to automated translation errors.

Civil rights

Reported: The report explicitly describes a violation of human and civil rights, specifically the denial of asylum.

Directly caused: The translation error directly led to the rejection of the refugee's asylum claim, violating her right to seek safe refuge.

Indirectly caused: N/A

Inferred additional harm: Numerous other asylum seekers may have had their legal rights compromised due to unverified automated translations used by immigration officers.

Epistemic

Reported: The report explicitly describes epistemic harm where the translation tool fabricated a false narrative by swapping pronouns.

Directly caused: The tool changed 'I' to 'we' in the written statement, creating a false discrepancy that misled the judge.

Indirectly caused: N/A

Inferred additional harm: N/A

People affected

• Occurrences reported: 1
• People reportedly harmed: 1
• People reportedly exposed: 1

Potential causes

Management

• Prioritizing Cost over Safety: Companies use neural translation to cut costs and boost productivity.

Technology

• Pronoun Translation Errors: Automated tool swapped 'I' pronouns to 'we' in the written statement.
• Low-Resource Language Lag: AI tools for Pashto and Dari lag behind dominant high-resource languages.
• Lack of Cultural Awareness: AI fails to translate metaphors, idioms, and regional colloquialisms.

Data Inputs

• Limited Online Texts: Fewer texts available online for training Pashto translation models.
• Missing Specialized Vocabulary: Lack of training data for specific terms like family relations and ranks.

Human Factors

• Over-reliance on Automation: Treating machine translation as the ultimate solution instead of a helper.

Process and Methods

• Lack of Human Verification: Failing to check machine-generated translations with human experts.

Regulatory Environment

• Permissive Agency Policies: Immigration agencies independently choose to opt in to automated tools.

Information quality

• Classification confidence: High
• Reason for confidence: The report provides clear, firsthand testimony of a specific translation failure that directly resulted in a denied asylum claim, as well as broader context on how these tools are deployed in immigration systems.
• Ambiguities identified: The specific brand of the automated translation tool used in the 2020 incident is not explicitly named, though Google Translate is discussed as a primary tool used by agencies.

An Afghan refugee's asylum claim was denied in the US due to an automated translation tool error that swapped pronouns. While highlighting systemic risks of using automated tools in high-stakes government workflows, the incident has minor national security implications, primarily affecting administrative decision-making and human rights processes.

• Overall national security impact: Minor
• Response level: Moderate
• Scope: Single nation
• Primary target: United States
• Alleged perpetrator: Unknown

Threat characteristics

• Imminence: Long-term. The reliance on faulty machine translation in administrative workflows is an ongoing, long-term systemic issue rather than an active, immediate crisis.
• Autonomy: Human-supervised. The AI system acted autonomously to translate text, but its output was ingested into a human-supervised judicial process where humans made the final decision.
• Novelty: Established threat. Errors in machine translation, especially for low-resource languages, are a well-established and known technical limitation.

Impact by dimension

• Physical security: Negligible. No physical threat, kinetic attacks, or critical infrastructure disruptions occurred during this incident.
• Information security: Negligible. The incident involves a translation error in an administrative asylum process, with no evidence of information warfare or intelligence compromise.
• Sovereignty: Minor. The incident highlights vulnerabilities in government decision-making processes (immigration adjudication) due to automated translation errors, though it does not threaten state authority.
• Economic security: Negligible. No economic warfare, financial system attacks, or strategic technology theft occurred.
• Societal stability: Minor. The incident represents a human rights concern and unequal treatment of asylum seekers due to language-processing failures, but has limited scale in terms of overall societal stability.