AI-generated summaries of asylum interviews used by Britain's Home Office failed 22% of first-stage expert spot checks after the tool was rolled out to all caseworkers, more than double the rate the department reported from its pilot, according to a report by digital-rights campaign group Open Rights Group (ORG) covered by The Register Tuesday. The findings, drawn from Freedom of Information (FOI) requests and ministerial correspondence, are a test case for any organization relying on human review to catch generative AI errors in high-stakes decisions.

What Happened

The Home Office uses two tools built on OpenAI's GPT-4o model, accessed through Microsoft's Azure AI Foundry. Asylum Case Summarisation (ACS) condenses interview transcripts for the officials who decide claims and was rolled out to all asylum caseworkers in April 2026. Asylum Policy Search (APS), a chat-style tool for searching country policy information, has been in use by all asylum decision-makers since July 2025, according to ORG's report, "The Use of ChatGPT in Asylum Decisions: Errors, Inaccuracies & Home Office Incompetence."

Home Office subject-matter experts checked 203 ACS summaries. At the first stage, 139 passed, 19 were flagged for further review and 45 failed. Of 56 summaries that went to a second-stage review, 30 failed. The department's own pilot evaluation, published in April 2025, had found 9% of summaries inaccurate or incomplete. ORG adds that a summary omitting material facts can still pass, because the scoring criterion for relevance is not one of the grounds for failure.

RelatedNew York City Comptroller Urges Anthropic to Delay IPO Over Safety and Governance Concerns

The checks covered a sliver of the work. Citing the tool's data protection impact assessment, ORG says about 17,515 cases are processed a month, or roughly 70,000 between April and July, of which 203 were checked, about 0.3%. A feedback inbox for caseworkers received seven issues over the same period, the report says.

For APS, ORG says ministers told Parliament in April and an MP in June that expert testing was continuing, but Border Security and Asylum Minister Anna Turley confirmed in August that it had ceased and that no summary of its findings was produced. Applicants are not specifically told AI has been used in their claims and cannot see or correct summaries of their own interviews, according to ministerial correspondence the report cites.

Why It Matters

The case shows a governance gap that reaches well beyond immigration. "Human in the loop" is the standard reassurance for generative AI used in decisions about people, whether in hiring, insurance claims or compliance reviews, but it means little if few outputs are checked and errors of omission are not counted as failures. Boards approving similar tools can ask what share of outputs is sampled, who reviews them and whether the people affected can see and correct what the AI wrote.

ORG also argues that ACS took over a summarizing step previously done by experienced technical specialists, removing a layer of scrutiny rather than merely saving time, and it quotes the department's own data protection assessment describing the tool as automating that step. Promised efficiency gains are in question as well: ORG's FOI data show average monthly initial asylum decisions since April 2026 down about 42% from the same period in 2025, and mean time from interview to decision up 29%. The Register cautioned that those figures do not prove AI caused the slowdown.

The government says humans remain responsible for asylum decisions, The Register reported, and the Home Office did not respond to the publication's questions. According to ORG, the department did not apply the government's AI Playbook, its guidance on safe AI use in the civil service, on the grounds that the pilots predated it, although both tools were fully rolled out after the Playbook's February 2025 publication.

What to Watch

  • The Home Office's response to ORG's call to suspend both tools, commission and publish an independent audit, and ask the Independent Chief Inspector of Borders and Immigration to inspect them. A minister wrote in June that the department did not plan an independent audit "at this stage."
  • ORG's complaint to the Information Commissioner's Office over withheld training materials, which the report says is under investigation.
  • Whether the Home Office publishes equality impact assessments for the tools, which a minister said it aimed to release in August.