UK Uses AI to Summarise Asylum Interviews: One in Five Checked Summaries Failed

# UK Uses AI to Summarise Asylum Interviews: One in Five Checked Summaries Failed Source: [https://www.ibtimes.sg/uk-uses-ai-summarise-asylum-interviews-one-five-checked-summaries-failed-95621](https://www.ibtimes.sg/uk-uses-ai-summarise-asylum-interviews-one-five-checked-summaries-failed-95621) Forty\-five of 203, that is, how many AI\-generated summaries of asylum interviews failed an initial expert check after the UK Home Office rolled out its summarization tool to every caseworker in April 2026, according to figures obtained by digital rights group Open Rights Group \(ORG\)\. The failure rate of about 22% is more than double the 9% recorded during the government's own pilot\. Britain's borders watchdog is now considering its inspection priorities for next year and has asked ORG for more information about the tools\. This is not a confirmed investigation, but it puts the Home Office's use of AI in asylum processing under closer scrutiny\. The concern is not that a machine formally decides who gets asylum\. The Home Office says it does not\. The question is whether caseworkers can make reliable decisions when the summaries in front of them contain errors\. ### Watchdog Seeks Information on Home Office AI Tools ### The Independent Chief Inspector of Borders and Immigration has asked ORG for more information as it plans its 2027 inspection topics, according to WIRED\. The request signals that the issue is under consideration, not that a formal investigation has begun\. The Home Office uses two tools: Asylum Case Summarisation \(ACS\), which condenses interview transcripts for decision\-makers, and Asylum Policy Search \(APS\), a chatbot\-style tool for country information\. Documents obtained by ORG show both run on[OpenAI's](https://www.ibtimes.sg/openai-ai-says-famous-computer-science-hypothesis-false-thats-just-one-three-discoveries-95571)GPT\-4 family through[Microsoft's](https://www.ibtimes.sg/microsoft-offers-buyouts-60000-workers-amid-ai-spending-layoff-concerns-95563)Foundry platform\. OpenAI told WIRED it plans to retire GPT\-4 in January\. A Home Office spokesperson said trained caseworkers make the decisions and that the tools "cannot be used in isolation and do not make decisions on claims\." ### 22% of Summaries Fail Initial Expert Checks ### ORG's six\-month investigation drew on 11 parliamentary questions, two ministerial letters and 10 public\-record requests\. The Home Office's pilot evaluation found that 9% of ACS summaries were inaccurate or incomplete\. ### After the April rollout, 45 of 203 summaries checked by experts failed the initial review\. Those checks covered roughly 0\.3% of an estimated 70,000 cases processed between April and July\. The Home Office said in its April 2025 evaluation that summaries deemed inaccurate or incomplete were removed from the pilot\. However, the government later stopped expert testing, despite previously telling MPs that checks were continuing, The Register reported\. The figures describe summaries, not final asylum decisions\. A failed summary check does not establish that an applicant was wrongly refused asylum\. The Home Office has cited efficiency gains from the tools\. Its pilot estimated that ACS saved about 23 minutes per case and APS about 37 minutes\. With the asylum backlog above 62,000 claims in September 2025, the government had a reason to seek ways to speed up processing\. Applicants Are Not Told When AI Summarises Their Interviews ORG says asylum applicants are not told when AI is used in their cases, leaving them with no clear way to check or challenge an[AI\-generated](https://www.ibtimes.sg/ai-reshaping-2026-midterms-137-million-spent-ai-generated-political-ads-95266)summary of their interview\. An interview may contain detailed accounts of persecution, threats or events that are difficult to explain\. If a summary omits a qualification or changes the meaning of an answer, a caseworker could begin reviewing the claim with an incomplete account\. Officials may have access to the original transcript, but the summary can shape how the case is first understood\. ORG commissioned an 84\-page legal opinion from barristers at Cloisters and Doughty Street\. The barristers argued that asylum\-seekers have a right to be told when AI is used to determine their claims, how it is used and to receive the AI\-generated summaries\. The opinion concluded that the Home Office's use of AI is likely to be unlawful\. This is a legal assessment commissioned by ORG, not a court ruling\. The UK's data regulator is separately investigating the Home Office's refusal to disclose information about how staff use the tools and how they were trained to work with them, WIRED reported\. ### Tribunal Judge Flags Possible AI Use in Asylum Refusal Letter In a July 14 ruling, an Upper Tribunal judge said a Home Office letter refusing a Moroccan woman's asylum claim bore hallmarks consistent with the use of artificial intelligence\. The letter cited a country information note on Morocco dated July 2021, which the judge wrote appeared never to have existed\. The judge did not conclude that AI had written the letter, describing that possibility as something that might become clear\. The available reporting also does not link the letter to ACS or APS\. The ruling nevertheless raises questions about the reliability of material used in asylum decisions, particularly when inaccurate or potentially fabricated information reaches a decision issued under human authority\. The inspectorate has yet to confirm whether the tools will feature in its 2027 inspection plan\. The data regulator's findings and any legal challenges based on ORG's commissioned opinion could provide further tests of the[Home Office's](https://www.ibtimes.sg/bigil-box-office-collection-here-how-much-vijays-film-has-collected-its-home-territory-34557)approach\. For now, the figures do not prove that AI has caused wrongful asylum refusals\. They show that 45 summaries in a small expert sample failed initial checks, while broader expert testing has stopped\. The Home Office says humans make the decisions, but the unresolved question is whether they have accurate information in front of them, and whether applicants can challenge the summaries that may shape how their cases are understood\.