The United Kingdom’s use of generative AI in asylum processing is facing renewed scrutiny after an investigation found that more than one in five interview summaries examined by experts failed initial accuracy checks. The Register reported on October 13 that the findings concern a Home Office system used in decisions that can determine whether a person is allowed to remain in the country.

The findings come from a 20-page investigation by the Open Rights Group, based on Freedom of Information requests and correspondence with ministers. The group examined two tools used by Home Office caseworkers: Asylum Case Summarisation, which produces summaries of asylum interviews, and Asylum Policy Search, a chat-style system for finding country-policy information. According to The Register, both use OpenAI’s GPT-4o through Microsoft’s Azure AI Foundry.

Interview transcript becomes a shorter summary with visible gaps and missing pieces.
A summary may contain accurate statements while still omitting details important to an asylum claim.

Figures obtained by the campaign group showed that 45 of 203 AI-generated interview summaries failed initial expert checks after the summarization system was introduced in April 2026. That is a 22 percent failure rate, compared with 9 percent during the pilot. A further 19 summaries were referred for additional review. Of 56 second-stage reviews recorded in the figures, 30 did not meet the required standard.

The reported errors included factual inaccuracies and distorted accounts of what asylum seekers said during interviews. The Open Rights Group also found that a summary could pass a check despite leaving out important details, provided the material it did include was accurate. That distinction matters because an apparently clean summary may still fail to preserve information relevant to a person’s claim.

The scale of quality control is another point of concern. The group estimated that Home Office experts checked only 0.3 percent of summaries produced between April and July, using the department’s projected monthly caseload of 17,515. The government says people remain responsible for asylum decisions, but the group questions whether caseworkers can consistently detect errors in AI-generated material when so few outputs receive expert examination.

A few case folders receive expert review while many others continue along a conveyor.
The investigation estimated that expert checks covered only a small fraction of the system’s generated summaries.

The available figures do not demonstrate that AI made asylum processing slower. They do, however, offer little support for promised productivity gains. The Register reported that the average monthly number of initial substantive decisions since April was about 42 percent lower than in the same period of 2025, while the average time from interview to decision rose 29 percent. The source explicitly noted that these comparisons do not prove causation.

Oversight questions extend to the separate policy-search tool. The Home Office confirmed in August that expert testing had stopped after previously telling members of Parliament that checks were continuing, according to the report. The Open Rights Group said no summary of the testing results was produced. Applicants are also not specifically informed when AI has been used in their cases and cannot inspect or correct the generated interview summaries.

The rights group is calling for both systems to be suspended and independently audited. It argues that the summarization tool may replace work formerly performed by technical specialists, removing an experienced layer of scrutiny rather than simply accelerating an existing task. The Home Office did not respond to questions from The Register. Until fuller auditing is available, the central uncertainty is not whether humans formally retain the final decision, but whether they receive a sufficiently accurate and complete account on which to base it.