Case Studies / UK Asylum System
AI in the Asylum System
A map of where AI now sits in UK asylum casework. The applicant's own-language testimony stays in audio. The decision record works from English.
Read: The Home Office Is Using AI on the Wrong DocumentWhy the personal layer matters
Machine translation can be accurate and still remove the personal, relational layer of language — tone, register, dialect, hesitation, idiom, rapport. People who used Google Translate described exactly this: the words were right, but it did not feel personal or connective.
That personal layer is not decoration. In an asylum hearing, it is the very thing a person is judged on — credibility is assessed on how someone speaks as much as what they say. So a translation that strips the personal layer strips the basis on which a person is believed.
This is why machine translation is inadequate for high-stakes settings, even when it is accurate: it removes the subtleties used to judge a person. And it is why human interpreting — which can carry that layer — must be done properly, because where it fails (wrong dialect, "good enough" standards), the same loss occurs.
Processing chain
From policy to appeal.
Original testimony - not yet recorded.
01 Home Office Policy Policy layer Original testimony - not yet recorded.
What happens
Policy tells caseworkers what evidence to ask for, how to read credibility, and how to grant or refuse protection.
Is AI involved?
No active AI tool is flagged here. This layer still matters. Later tools draw from policy and casework habits.
Known limitations
Policy treats the English record as the applicant's account. It rarely marks it as an interpreted document.
02 ILSU / Interpreter Booking TheBigWord contract Procurement layer Interpreter assigned - no qualification verified.
What happens
The Home Office books interpreters through language-service procurement. That booking happens before testimony becomes an English record.
Is AI involved?
No active AI tool is flagged here. The risk sits in procurement. Availability decides whose English enters the record.
Known limitations
Inspection evidence found interpreter shortages, cancelled interviews, dialect mismatch, and quality assurance gaps.
Source: ICIBI, An inspection of asylum casework, June to October 2023
03 Asylum Substantive Interview Testimony layer Applicant speaks. Interpreter renders in English.
What happens
The applicant speaks through an interpreter. The interview becomes the main evidence used by the Home Office.
Is AI involved?
No active AI tool is flagged at the moment of speech. This is where the later record begins.
Known limitations
Audio captures the room, including the applicant's own language. The transcript records only English: the interpreter's rendering, attributed to the applicant.
Source: Right to Remain Toolkit, Asylum Substantive Interview; Vogl, Refugee credibility assessment and the vanishing interpreter
04 Asylum Case Summarisation Tool ACS AI active AI summarises the English rendering.
What happens
ACS summarises asylum interview transcripts for caseworkers. It reads the English record produced through interpretation.
Is AI involved?
Yes. AI is active here.
Known limitations
ACS works from the English transcript. It summarises the interpreted version. Errors, omissions, and flattening can all enter the summary.
Source: Home Office, Evaluation of AI trials in the asylum decision-making process; Open Rights Group legal opinion press release
05 Asylum Policy Search Tool APS AI active Policy search runs on AI-summarised CPINs.
What happens
APS helps users find policy material for asylum casework.
Is AI involved?
Yes. AI is active here.
Known limitations
Search tools decide what gets surfaced, missed, or prioritised. Applicants cannot see whether an AI search shaped their case.
Source: Electronic Immigration Network, Home Office to expand AI use in asylum decision-making; Open Rights Group legal opinion press release
06 Case Decision Decision layer Decision made on AI summary of interpreted transcript.
What happens
A caseworker grants or refuses using the record, policy guidance, transcript, summaries, and case materials.
Is AI involved?
No final automated asylum decision is flagged here. The risk is earlier: AI can shape the material a human reads.
Known limitations
Human review cannot repair a record built only from English if the applicant's own-language testimony stays untranscribed.
07 Appeal Challenge layer Appeal - original testimony still not available as evidence.
What happens
If refused, the applicant can appeal. Lawyers and judges work with the record that survived.
Is AI involved?
No active Home Office AI tool is flagged here. Appeals inherit the gaps.
Known limitations
Without an original-language transcript, an appeal cannot compare the English record with what was said. It has to go back to raw audio.
The Gap
Final verdict.
By decision stage, the original testimony is usually raw audio. Everything else is a rendering of a rendering.
What this demands
The three-layer problem exists because the system never had to preserve what was said. Reform 2a keeps original-language transcripts. Reform 2b makes AI work from them.
Currently being presented to a member of the Home Affairs Select Committee in July 2026.