TL;DR

The independent watchdog for borders and immigration, ICIBI, is considering whether to look at how the Home Office uses generative AI in asylum casework, WIRED reported on 27 October. The inspectorate has asked Open Rights Group (ORG) for more information while it plans next year’s inspection topics. The ICIBI would not confirm what is under consideration.

What is being looked at

Two tools are in question. Asylum Case Summarisation (ACS) condenses transcripts of interviews with applicants, and Asylum Policy Search (APS) finds and summarises country policy material. Documents obtained by ORG show both run on OpenAI’s GPT-4, released in 2023, through Microsoft Foundry. The Home Office began piloting them in late 2024 as Labour worked through roughly 100,000 outstanding asylum cases, and its own estimate is that together they save about an hour per case.

According to ORG campaigner Sara Alsherif, ICIBI’s request follows a six-month ORG investigation built on parliamentary questions, ministerial letters and public records requests. ORG says the share of ACS summaries judged unfit for use is well above the 9% error rate published in the pilot evaluation, and that APS output has no quality control at all. ICIBI told WIRED it was aware of concerns and pointed to its own finding last year that one tool’s “effectiveness and reliability were varied”.

Transparency gaps

The UK’s AI Playbook says departments should list algorithmic tools used in decisions on a public repository. Neither tool appears there. Applicants are not told when the tools are used on their case, and Cloisters lawyers, in an opinion ORG commissioned, say that being unable to challenge an inaccurate AI summary may breach UK data rights law. Separately, the Information Commission is looking into why the Home Office declined to explain to ORG how its staff use, and were trained on, two of its AI tools.

The Home Office says trained caseworkers make every decision and the tools do not decide claims. OpenAI told WIRED it will retire GPT-4 in January.

This builds on our earlier coverage of expert spot checks on the summaries and ORG’s wider report on Home Office AI use.

Looking Forward

An inspection would put an official watchdog’s findings alongside the evidence campaigners have pieced together through FOI requests. In our view, the retirement of GPT-4 adds a practical question: if the tools stay in use, the department may have to move them to a newer model, and any re-evaluation is a natural point to publish them on the repository. The Public Law Project says it has found 11 high-risk AI systems across government that are missing from that register.