The UK’s borders watchdog is examining how the Home Office uses generative AI to process asylum claims. The tools in question help caseworkers summarise interviews and search policy, in cases that can decide whether someone gets to stay in the country.
The Independent Chief Inspector of Borders and Immigration (ICIBI) has asked digital rights nonprofit Open Rights Group (ORG) to share more about the tools. The inspectorate is currently drawing up the list of topics it will inspect next year, ORG campaigner Sara Alsherif told WIRED.
Two chatbots inside the asylum system
The Home Office has deployed two generative AI tools in asylum casework. Both run on OpenAI’s GPT-4o model and are accessed through Microsoft Azure.
The first is the Asylum Policy Search (APS) tool. It gives staff a chat interface for looking up country-specific policy. APS went live for all decision-makers in July 2025.
The second is the Asylum Case Summarisation (ACS) tool. It condenses interview transcripts into summaries for caseworkers. ACS reached caseworkers in April 2026.
The accuracy numbers moved in the wrong direction
During the Home Office’s own pilot evaluation, 9% of ACS summaries were found to be inaccurate or incomplete.
After the tool rolled out, the figure got worse. Post-rollout checks on 203 ACS summaries found a 22% failure rate.
Oversight is also thin. Only 0.3% of the approximately 7,000 summaries the tool generates each month were reviewed by subject matter experts.
Asylum applicants are not told when AI tools are used on their cases. They also cannot view the summaries generated about them, or correct them.
ORG obtained a legal opinion on that point. It concluded that the failure to notify applicants could render the tools unlawful.
How the findings surfaced
ORG spent six months digging into the tools, and the Home Office withheld details along the way.
The group’s effort involved 11 parliamentary questions, two ministerial letters, and 10 public records requests.
Based on what it found, ORG argues the deployment falls short of established guidance on using AI in sensitive roles. It has called for the tools to be suspended immediately and for independent audits.
The Home Office maintains that humans remain in charge. In a statement, it said:
“Asylum decisions are made by trained caseworkers, with AI tools used only to assist in navigating interview records and policy information.”
Why the ‘human in the loop’ defence is under pressure
With a 22% failure rate in post-rollout checks and expert review of just 0.3% of output, the safety net looks more like a few loose threads. And because applicants cannot see the summaries, the person with the most knowledge of what was actually said has no way to flag a mistake.
What this means
For the Home Office, the immediate risk is legal and reputational. ORG’s legal opinion suggests the lack of applicant notification could make the tools unlawful.
The ICIBI’s interest raises the stakes further. The inspectorate is only weighing its topics for next year, so nothing is guaranteed. But an inspection would give an independent body formal access to information that ORG had to pry out through months of parliamentary questions and records requests.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.