The UK’s border watchdog is considering an inspection of the Home Office’s use of generative AI in asylum casework. The tools help staff review interview records and country information, but digital rights groups say questions remain about their accuracy, oversight and transparency.
Tools built to help caseworkers review claims
The Home Office uses two systems: Asylum Case Summarisation (ACS) and Asylum Policy Search (APS). Documents obtained by Open Rights Group (ORG) say they are built on OpenAI’s 2023 GPT-4 large language model and Microsoft Foundry.
ACS summarizes interviews in which applicants describe persecution and explain why they fear returning to their country. APS searches and summarizes policy information about countries of origin, which can help caseworkers assess whether an applicant faces a genuine risk of harm.
The Home Office began piloting the tools in late 2024 while the Labour government sought to address a backlog of around 100,000 asylum cases and handle tens of thousands of new applications the UK receives each year. Government estimates from the pilot said using both tools could save around an hour per case.
The department says trained caseworkers make asylum decisions and that AI tools assist with navigating records and policy information. A spokesperson said the tools cannot be used alone to decide claims, and that the Home Office monitors their performance alongside its quality-assurance processes.
Accuracy and oversight are under scrutiny
ORG’s six-month investigation found that the share of ACS summaries judged not fit for use was significantly higher than the published 9 percent error rate calculated in the pilot evaluation. The group said these summaries were often unsuitable because of factual errors or misrepresentations. It also found no quality control for APS summaries.
Those findings matter because the tools handle information that may shape how a caseworker understands an applicant’s account and country conditions. A summary that leaves out or distorts details could affect the material a human decision-maker reviews, even if the final decision remains with that person.
The concern is especially acute in cases where an applicant fears persecution if returned. In July, a UK judge accused the Home Office of citing an “AI hallucinated” document when refusing an asylum claim by a Moroccan woman and her child on the basis that her home country would be safe for her to return to.
Campaigners and the Public Law Project have also pointed to concerns about “cultural bias” in GPT models. They warn that language barriers or missing cultural nuances in summaries could contribute to discriminatory decisions. An OpenAI spokesperson told WIRED that the company is working to improve model accuracy and said it will be retiring GPT-4 in January.
Watchdog and public visibility questions
The Independent Chief Inspector of Borders and Immigration (ICIBI), which monitors the Home Office’s management of borders, immigration and customs, has asked ORG for information as it plans topics to inspect next year. The watchdog said it could not confirm which topics are under active consideration before publishing its annual inspection priorities, but it acknowledged concerns about AI use and cited last year’s finding that one tool’s “effectiveness and reliability were varied.”
ORG’s inquiry involved 11 parliamentary questions, two ministerial letters and 10 public records requests seeking details the Home Office has withheld. The organization and Public Law Project have sent findings to the Information Commission, which is investigating the Home Office’s refusal to explain how staff use the tools and how they were trained to use them.
Transparency is another unresolved issue. The UK’s “AI Playbook” says departments must be open about algorithmic tools used in decision-making and make information about them public, including by listing them in a repository. The Home Office’s case summarization and policy search tools are not listed there, and asylum applicants are not told when the systems are used in their cases.
ORG campaigner Sara Alsherif says migrant organizations and legal practitioners often do not know the tools are in use, making it harder to understand or challenge their role. Lawyers at Cloisters wrote in an independent legal opinion commissioned by ORG that preventing someone from contesting an incorrect AI summary of their personal information may breach UK data rights laws.
Disclosure shapes the ability to challenge errors
The debate is not only about whether AI can save staff time. It is also about whether applicants and their representatives can see when a tool has influenced the material considered in a case, identify an error and raise it before a decision is made.
Public Law Project senior research fellow Caroline Selman said ACS and APS are two of 11 high-risk AI systems her team has found in government that are missing from the public repository. She warned there may be other systems in use that researchers, and the people affected by them, do not know about.
The Home Office says AI does not replace or automate human decision-makers. Watchdog scrutiny, public disclosure and the ability to challenge errors will help clarify how that assurance works in practice for people seeking asylum.