Home Office AI tool for asylum decisions under fire

Government run trial finds errors

An artificial intelligence tool developed by the Home Office to help clear the UK’s asylum backlog is being prepared for rollout despite producing “serious errors” during a government-run trial.

According to official documents obtained via Freedom of Information laws by campaign group Foxglove and shared with The i Paper, nearly one in 10 (9%) of AI-generated summaries of asylum interviews were so inaccurate or incomplete that they had to be excluded from the pilot.

The AI, known as the Asylum Case Summarisation tool, uses a GPT-4-based model to analyse and summarise asylum interview transcripts. The government claims it will help speed up the processing of asylum applications and appeals, with more than 140,000 currently outstanding.

However, legal experts, MPs and refugee advocacy groups have warned that hasty adoption of the tool could result in harm to vulnerable applicants, particularly in cases involving life-or-death decisions.

Tony Vaughan KC, Labour MP and chair of the cross-party parliamentary group on refugee issues, said AI “can never be a substitute for reading all the documents” and called for more robust oversight before rollout.

Growing concerns over accuracy and transparency

While the government has touted the ACS tool as a success highlighting a potential 30% reduction in time spent summarising interviews, its own internal evaluation reveals deeper concerns.

The Home Office report noted that 23% of caseworkers who tested the AI were “not fully confident” in the summaries it generated.

The ACS tool was developed internally by the Home Office using OpenAI’s GPT-4 model and is scheduled for full rollout by January 2026. A senior civil servant confirmed the deployment will take place “over several months”, but no start date has been made public.

The tool is designed to automate the labour-intensive task of summarising long asylum interviews, with the summaries intended to assist, not replace, human caseworkers in deciding the outcome of claims.

However, the known limitations of LLMs including hallucinations, where AI generates plausible but false information have raised red flags. Foxglove, which campaigns for fair technology in public services, has urged a halt to the rollout.

Tim Squirrell, the group’s head of strategy, said: “Asylum decisions can be a matter of life or death. You probably wouldn’t want to trust your life to a chatbot making such serious errors that its work has to be thrown out nearly 10 per cent of the time.

“We’re calling for the rollout of this dodgy bot to be put on hold until the Government fesses up about how it intends to deal with this.”

Foxglove also requested the tool’s Equality Impact Assessment under FoI laws, but the Home Office declined to provide it.

A broader AI strategy under scrutiny

The ACS tool is not the only AI system being explored by the Home Office in the asylum process. A second tool, known as Asylum Policy Search, is under development to help caseworkers locate relevant policy or country guidance.

Yet uptake has been lukewarm. Just 54% of staff involved in the trial said they would continue using the tool, with some reporting that they “did not see the benefit” compared with current search systems.

This is not the first time the Home Office has faced controversy over its use of AI in immigration processes. In 2020, it scrapped a visa algorithm accused of racial bias, which critics said created a “fast lane” for applicants from majority-white countries. Earlier this year, it also announced plans to use AI to assess the ages of asylum seekers claiming to be children, despite a watchdog warning that mistakes would be “inevitable”.

A Home Office spokesperson said: “This new AI technology is helping to speed up asylum processing, as we continue to implement our Plan for Change to restore order to the asylum system, clear the backlog and end the use of hotels by the end of this Parliament.

“The new tech will support caseworkers to make accurate, evidence-based decisions whilst reducing the time taken to do resource-intensive administrative tasks in support of that process.”