The Home Office is using ChatGPT-based software to summarise asylum interviews, and the people whose testimony is being summarised are not told. In production, 22% of those summaries failed a first quality check. Only 203 of about 70,000 were reviewed at all, according to a six-month investigation by the Open Rights Group (ORG), a British digital rights organisation.
The people affected are overwhelmingly migrants and refugees from Africa, the Caribbean and Asia, the same communities the Home Office’s “hostile environment” has targeted since the Windrush scandal. The report, “Automated Discrimination: Unpacking the Digital Hostile Environment”, argues that AI is now being added to that machinery with almost no public disclosure.
Two tools sit at the centre of it. The Asylum Case Summarisation tool condenses interview transcripts for caseworkers. The Asylum Policy Search tool lets them query country policy information through a chat interface. Both run on GPT-4o. The search tool was rolled out in July 2025 and the summarisation tool in April 2026. Neither appears on the government’s own AI register.
The Home Office’s pilot found 9% of summaries so flawed they had to be removed. In live use between April and July 2026, ORG found, 22% failed at first review and 15% still failed after a second check. Minister Anna Turley confirmed in August that expert testing of the search tool had stopped entirely.
A legal opinion ORG commissioned in March concluded that the tools are likely unlawful, citing procedural fairness, data protection and equality law. Barristers warned of “life-changing decisions based on material errors of fact”, including omitted evidence or invented detail.
Beyond the interview room
The report goes further than asylum interviews. It lists AI used or tested to estimate asylum seekers’ ages, flag suspected sham marriages, search the phones of new arrivals, prioritise cases through a tool called Ipic, and watch borders with facial recognition and drones. ORG says this is happening “in near-total secrecy, with officials routinely invoking national security” to refuse freedom of information requests.
Age assessment shows how the risk falls unevenly. Human Rights Watch and 62 other groups have urged the Home Office to drop facial age estimation for people claiming to be under 18. The Home Office has acknowledged that top systems carry “an error margin of around 2.5 years” at that boundary and that performance “can vary depending on ethnicity”. A child wrongly judged an adult can be housed with adults and lose child protections. The department has declined to publish its equality and data protection impact assessments.
Can anyone challenge it?
Hardly, because a person who is never told AI summarised their interview has nothing to contest. The legal opinion says applicants should be told, and should get access to the AI-generated material.
The Home Office rejects the picture of machines deciding claims. A spokesperson said trained caseworkers make every decision, and the tools “cannot be used in isolation and do not make decisions on claims”. The department says caseworkers must read the underlying documents. ORG points to the Home Office’s own estimate of 23 minutes saved per transcript as reason to doubt that happens every time.
ORG wants both tools suspended, protections for affected applicants and an independent audit of accuracy in live use. The Independent Chief Inspector of Borders and Immigration is weighing an inspection for 2027, and the Information Commissioner’s Office is investigating whether the Home Office improperly withheld staff training material.
Have your say