Q: What is the Asylum Case Summarisation (ACS) tool and how does it work? A: The ACS tool, introduced in April 2026 and based on GPT-4o, reads interview transcripts and produces automated case summaries that can be up to 50 pages long before they are reviewed by decision makers. Ministers say it aims to speed up decisions and saves about 23 minutes per case, and a separate Asylum Policy Search (APS) tool helps with country-policy searches. Q: What kinds of errors have been found in ACS summaries during testing? A: Home Office asylum AI errors identified in testing include factual mistakes (wrong names, dates or places), contradictions, misstatements of quotes and omissions of material facts. In dip-sample testing 22% of 203 summaries failed in the first stage, 15% failed in a later review, and 9% of pilot summaries were removed for inaccuracy or missing information. Q: How many ACS summaries were checked by Home Office experts during the rollout? A: Home Office subject-matter experts dip-sampled 203 summaries out of an estimated 70,000 generated between April and July 2026, representing roughly 0.3% of AI-generated summaries. The DPIA also notes an average of about 17,515 asylum cases processed per month, which highlights the small scale of spot-checking during rollout. Q: Are asylum applicants informed when AI-assisted summaries are used in their cases? A: Ministers have confirmed that asylum applicants are not told when AI has been used in the consideration of their claims. Campaigners such as the Open Rights Group say this lack of disclosure reduces transparency and can remove a layer of expert review. Q: What quick red flags should reviewers look for to spot risky AI-written summaries? A: To detect Home Office asylum AI errors quickly, reviewers should check for misspelled names and identifiers, timeline drift, vague or conditional wording around core claims, truncated quotes and omitted high-impact details such as persecution events or medical notes. Simple checks include a keyword sweep for risk terms, comparing quoted lines to the transcript and using a brief timeline table to verify chronology. Q: What practical safeguards does the article recommend to reduce harm without halting AI tools? A: To mitigate Home Office asylum AI errors the article recommends mandatory disclosure to applicants, documented human-in-the-loop checks, adding “material omissions” as a formal fail criterion and requiring direct transcript citations or policy references for credibility-affecting claims. It also advises risk-based sampling that scales beyond current spot checks, an error taxonomy with published performance reports, an escalation path to technical specialists and independent or external audits. Q: What steps can applicants and advocates take if they suspect errors in an AI-generated summary? A: Applicants and advocates should request the AI summary, the full interview transcript and the list of sources used, then submit a short errata sheet noting page and line references for each correction. They should flag omitted persecution events or medical details, verify policy quotes against the original guidance and keep a dated audit trail of disclosure and correction requests for appeals. Q: Are AI summaries included in the Home Office’s monthly quality assurance checks or planned for independent audit? A: Ministers have confirmed that AI-generated summaries are not considered during monthly quality-assurance checks, and the Home Office has told MPs it has no plans for an independent audit of the ACS. The department says trained decision makers make all asylum verdicts and that AI tools only support caseworkers, while spot-testing results have mainly been discussed verbally.