90 incidents on record · 2026 Headlights Incident reports by Ellie Harris · Melbourne
10 new this week Library last updated 13 July 2026
← The incident library
HD-INC-070
Immigration · United States · 2023 · Automated-decision harm

US immigration authorities leaned on machine translation for asylum cases, and errors as small as a pronoun reportedly led to rejections

By Ellie Harris · Filed Reported through 2023

Alleged: US Department of Homeland Security (deployer); Google Translate, Lionbridge, TransPerfect (tools and contractors, as reported) developed or deployed the AI system implicated in this incident. Details are drawn from public reports; parties are presumed innocent of any wrongdoing not established by an official finding.

US immigration authorities leaned on machine translation for asylum cases, and errors as small as a pronoun reportedly led to rejections

What happened

Through 2023, reporting by The Guardian, Rest of World, and Context documented US immigration authorities relying on machine translation in asylum processing, from Customs and Border Protection’s app-based tools to, per The Guardian, instructions for officials to use Google Translate when vetting refugee applications, alongside contracts with translation firms including Lionbridge and TransPerfect. The tools were being applied to the highest-stakes documents in the system: affidavits, interview records, and applications where a single wrong word can read as an inconsistency, and inconsistency can mean refusal.

The documented errors were small and consequential. Respond Crisis Translation, a translator network whose casework anchored the Guardian reporting, described an Afghan woman’s asylum affidavit in which a machine rendered every “I” as “we”; the group says her application was then rejected, in its account because the claim appeared to be filed on behalf of more than one person. In another case the group described, a domestic violence survivor referred to her abuser as “mi jefe”, a colloquialism for her father; the tool translated it as “my boss”, and the group says she was initially denied. Rest of World reported that machine translations of Pashto and Dari were riddled with errors, contributing to at least one rejected Afghan claim. The group says several of these cases were won only after human translators redid the work. The agencies have made no formal admission that translation errors caused wrongful denials, and no regulator has ruled on the practice.

What an auditable version would have shown

A refusal built on a mistranslation looks, in the file, like a refusal built on the applicant’s own words. That is what makes this failure mode so hard to appeal: the error wears the applicant’s voice. An auditable version binds every machine-translated passage in a case file to a signed record of what produced it: the tool and version, the source text, the output, and whether any qualified human reviewed it before it entered the decision. When an adjudicator cites an inconsistency, the record shows whether the inconsistency belongs to the applicant or to the software, which is the difference between a credibility finding and a transcription error.

Where the gap was

Machine output entered consequential legal decisions as if it were verified fact. A VerificationGate routes high-stakes translations to a trusted check before they can support a decision: a qualified human reviewer for the language pair, flagged confidence on low-resource languages, a block on unreviewed output in credibility determinations. A ConductRecord makes the tool’s role visible in the file itself, so an appeal can target the actual source of an error instead of arguing against a machine’s words attributed to a human.

What governance should have looked like

The rule for automated language in high-stakes process is the same as for any model output: it is a draft until something trustworthy has verified it, and its provenance must travel with it. A system that lets unreviewed machine translation stand as an applicant’s testimony has quietly delegated a life-altering credibility judgment to a tool that was never designed for it, and has left no trace of the delegation. The record should always be able to answer one question: whose words are these?

The reference implementation of VerificationGate and ConductRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.

Sources

The mailing list

Fresh incident reports every week. One email to match.

We add new incidents to the library regularly, and send a single short email each week with what's new. The library stays free and open; this is just how you keep up with it.

No tracking. Unsubscribe in one click.

The record

An auditable system would have produced a signed, tamper-evident record the moment this happened: what the system did, the version that did it, the basis it acted on, and the action taken, and US Department of Homeland Security (deployer); Google Translate, Lionbridge, TransPerfect (tools and contractors, as reported) could have produced it on demand.

This is the record the system as deployed did not produce in a signed, auditable form.

What this teaches
Capture what happened when it happens
What the system did, the version that did it, the basis it acted on, and the action taken, recorded at the moment, not reconstructed after.
Sign it, so no one has to trust the record-keeper
A tamper-evident entry. Edit it later and the signature breaks. The record does not ask for the benefit of the doubt.
Make it verifiable by anyone
A court, a regulator, a customer's lawyer can check the record themselves, without taking the company, or us, at our word.

Headlights summarises publicly reported AI incidents. All summaries are independently written, attributed to their original sources, and intended for research and educational purposes. Allegations are identified as such until established through official findings.

Last reviewed June 2026. This report is based on the sources listed above and reflects information available at the time of review; later developments may not be captured. Where a person is described as charged with or alleged to have done something, that allegation is unproven unless a conviction or a court or regulatory finding is stated. Headlights publishes journalism and commentary, not legal advice.

Want to write back?

Direct to my inbox.

ellie@useheadlights.com →