Open Rights Group finds 22% of checked Home Office asylum AI summaries failed accuracy tests

The UK Home Office promised faster asylum decisions with AI. A new Open Rights Group investigation instead finds that 45 of 203 checked interview summaries failed initial expert accuracy checks after April’s rollout. The group is calling for both asylum AI tools to be suspended and independently audited, challenging the safeguards around information used in asylum decisions.

The 20-page investigation draws on Freedom of Information requests and correspondence with ministers, according to The Register. It examines two tools used by Home Office caseworkers, both powered by OpenAI’s GPT-4o through Microsoft’s Azure AI Foundry. Their roles are different:

• Asylum Case Summarisation produces summaries of asylum interviews. The investigation’s accuracy figures concern these summaries, which condense what applicants told officials.
• Asylum Policy Search is a chat-style tool that helps officials search country policy information. The investigation raises a separate testing concern about this system.

Accurate details can still leave an incomplete account

Figures obtained by Open Rights Group put the initial failure rate at 22%, compared with a 9% rejection rate during the original pilot. Those percentages describe summaries subjected to checks, not all summaries produced by the system. Another 19 summaries were flagged for further review.

The figures also record 56 second-stage reviews, of which 30 failed to meet the required standard. These are a separate review stage, not another failure percentage to add to the initial result. Errors can include factual inaccuracies and misrepresentations of what applicants said in their interviews.

Passing a check did not necessarily mean the summary was complete. Open Rights Group found that summaries could pass even when important details were missing, provided the information included was accurate. That distinction puts the adequacy of the checking standard alongside the errors it caught.

Who can catch—and correct—the mistakes?

Open Rights Group estimates that experts checked only 0.3% of summaries between April and July. That estimate uses the department’s projected monthly caseload of 17,515; it is not a direct count of every summary produced. It raises a coverage concern distinct from how well individual summaries performed.

For Asylum Policy Search, the Home Office confirmed in August that expert testing had stopped, despite previously telling MPs that checks were continuing. Open Rights Group says no summary of the testing results was produced. The concern here is continuing scrutiny of the search tool, rather than the interview-summary failure rate.

The government maintains that humans remain responsible for asylum decisions. But the department’s data protection assessment says the summary tool automates work previously done by technical specialists. Those specialists reviewed and summarized cases before passing them to decision-makers. Open Rights Group warns that automation could remove an experienced layer of scrutiny.

Applicants themselves cannot inspect or correct the AI-generated summaries of their interviews. They are also not specifically told when AI has been used in their cases, The Register reports. That leaves the people whose accounts are being condensed outside the summary-checking process.

The promised speed gains remain unproven

Open Rights Group also challenges the efficiency case. It says the average monthly number of initial substantive asylum decisions since April 2026 fell about 42% against the same period in 2025. Meanwhile, the mean time from interview to decision rose 29%.

Those changes do not establish that AI caused the slowdown. They also do not demonstrate the faster processing ministers promised. The group’s proposed response is suspension and independent audits of both tools. The Home Office did not respond to The Register’s questions.