AI tool used in thousands of asylum claims failing basic accuracy tests

The Home Office rolled out the use of artificial intelligence to summarise interview transcripts with asylum seekers in April to speed up decision-making on applications in a push to slash a huge backlog.

MPs and charities warned that accuracy should not be sacrificed for speed in these “life and death” decisions.

Labour has made significant progress in bringing down the number of people waiting on an initial asylum decision, with 40,168 people waiting in June this year – down 56 per cent on 2025. However, the appeals backlog has ballooned, with 90,741 cases outstanding as of June 2026 – up 50 per cent.

Data detailing the use of the Asylum Case Summarisation (ACS) AI tool, from April to the end of August, shows that 259 checks have been carried out by ‘Subject Matter Experts’ on the summaries. Of these, 203 summaries were reviewed once, and 56 were reviewed a second time.

Of the 203 primary reviews, 45 (22 per cent) failed accuracy standards, and 19 were flagged as needing further review. Of the 56 secondary reviews, 30 (54 per cent) were deemed to be sub-standard, information obtained by the Open Rights Group found.

Summaries fail when they include “multiple factual errors or misrepresentations” or when they are “largely factually inconsistent” with what the asylum seeker actually said in their interview. The summaries also fail if they are deemed not useful to the caseworker.

A separate internal Home Office impact assessment from October 2025, disclosed to the campaign group under freedom of information laws and shared with The Independent, predicted before the roll-out that on average 17,515 asylum cases would be processed every month using the AI tool.

Asylum seekers are not told when AI is used on their claims, although the government has now made the ACS tool available to all decision-makers. It was initially used as part of a pilot project, along with another AI tool to search asylum policy, between May and December 2024.

The ACS system uses OpenAi’s GPT-4 model. Government data published from the pilot testing claimed that the ACS tool saved asylum decision makers 23 minutes per case on average - pointing to the “high cognitive load” of reading lengthy interview transcripts which can often be up to 50 pages long.

However, Home Office ministers, in correspondence with MPs, insisted that the AI summary “does not replace any part of the asylum decision-making process”, and that “decision-makers must continue to read all documents related to the claim”.

Decision-makers can also use another AI programme, called the Asylum Policy Search (APS) tool, which can search through Home Office information about countries of origin and summarise policy regarding that country.

Subject matter expert checks of the APS have now stopped but other periodic expert testing continues, the Home Office said.

Bell Ribeiro-Addy, Labour MP for Clapham and Brixton Hill, said: “This data shows that human oversight of the Asylum Case Summarisation (ACS) tool is worryingly limited and inaccuracy far too common in the few cases where its results have been checked.”

She added: “We need far more transparency about the role of AI in asylum decision-making”, saying the Home Office should “come clean about how many errors are slipping through unchecked and what it is doing to fix this”.

LibDem shadow attorney general, Ben Maguire MP, said: “Using AI to completely replace genuine human scrutiny when processing sensitive asylum cases risks miscarriages of justice, costly legal appeals and administrative chaos.

“The government’s reliance on this failing AI process may also put the UK in breach of international law to provide asylum applicants with a fair hearing.”

He warned that, while the government needed to slash the asylum backlog after it built up due to Conservative policy, “it cannot cut corners on basic legal standards to do so”.

Sara Alsherif, at the Open Rights Group, said there was a “lack of transparency about how AI tools are being used”. She added: “It has taken eleven parliamentary questions, two ministerial letters and eight FOI requests to uncover just some information about how AI is being used in the asylum system. There are still many contradictions and crucial information remains unpublished.”

Nick Beales, at the charity Refuge and Migrant Justice, said new technology should not “come at the expense of robust and reasoned decision-making”.

He added: “Asylum interviews are delicate processes, with vulnerable people having to recount traumatic events, often in a second language or via interpreters. Distrust or suspicion of those in positions of authority is also common, which can affect a person’s willingness or ability to disclose sensitive matters.

“We are therefore concerned about flawed interview records being fed into AI software and producing unreliable decisions that do not account for the real-life circumstances in which these interviews take place.”

Mr Beales added that ministers should “ensure that AI is never used as a substitute for human review of these life and death decisions.”

A Home Office spokesperson said: “Asylum decisions are made by trained decision-makers, not artificial intelligence. Our AI tools support caseworkers by helping them navigate interview records and policy information, but they cannot be used in isolation and do not make decisions on claims.

“We regularly monitor the performance of these tools alongside our established quality-assurance processes. All asylum claims continue to be considered on their individual merits by trained decision-makers.”