The Home Office is using AI to summarise asylum interview transcripts. This is happening now. The Asylum Case Summarisation tool, built with ChatGPT-4, turns an interview record into a shorter document for caseworkers. A March 2026 legal opinion by three senior barristers found this is likely unlawful, partly because asylum seekers are not told.
The legal opinion misses one problem. The issue is the document the AI reads.
What the tools do
The AI does not see what the applicant said
In an asylum interview, an interpreter renders the applicant's words into English. The transcript records that English. It does not record what the applicant said in their own language. The original words stay as unanalysed audio.
The AI summarises the interpreter's English. It does not summarise the testimony itself.
By the time a caseworker reads the AI summary, the testimony has crossed three layers:
- The applicant speaks in their own language
- The interpreter renders it in English, making choices, omitting details, simplifying phrases
- The AI summarises that English, compressing again
The original is not checked against the later layers. The decision is made on a summary of an interpretation.
This is about the document
The Home Office will use AI. Assertion will not stop that. Tools that save time can have a place in an under-resourced system.
If AI summarises an interview, it should summarise what the applicant actually said.
Right now it cannot. The original-language testimony does not exist as readable text anywhere in the system.
What would change this
One reform closes the gap: transcribe and retain the original-language testimony for every asylum interview.
Keep that transcript beside the English record. Then AI can work from testimony. Errors can be checked. Appeals can compare what was said with what English recorded.
This does not remove AI. It gives AI the right document. That is a basic standard for a system deciding people's lives.
The Home Office's own inspectors found in 2023 that staff believed 40 percent of cancelled asylum interviews were due to interpreter issues. The Home Office had no data to confirm or deny it because nobody collected it. Original-language transcripts would show what the applicant said, what the interpreter rendered, what the AI summarised, and what the caseworker decided.
The three-layer problem
By decision stage, the original testimony is three layers away from the record. Each layer changes it. Nobody can check the change because the earlier layer was not preserved.
AI did not create the failure. The system already refused to preserve what was said. AI just added another layer.
The case for original-language transcription is simple: if you decide someone's life, work from what they said.
The Home Office is working from a summary of a rendering. Three steps from the truth. Give the AI the right document.
The full reform is in the policy briefing.
See also the AI Topology Map and the ICIBI inspection analysis.
Mohammad Shehadeh founded Translational Justice and wrote The Manifesto of Translational Justice (2026).