Imagine that you have had to flee your home because staying would mean death. You make a dangerous journey to the United Kingdom, ask for asylum, and are told to sit in a small room and tell your story to an official. You describe the threats, the violence, the fear that followed you across borders. Everything depends on being believed. Now imagine that, before a caseworker ever reads your testimony, a machine has already decided what in your story is important and what can be ignored. This is not a hypothetical scenario. It is the reality for asylum seekers in Britain today, where the Home Office has quietly introduced generative artificial intelligence into the processing of asylum claims. The tools, known as the Asylum Case Summarisation and Asylum Policy Search systems, are built on OpenAI’s GPT-4 large language model and Microsoft Foundry. The first tool summarizes interview transcripts; the second helps caseworkers find and summarize official policy information about applicants’ home countries. They were piloted in late 2024, as the Labour government tried to clear a backlog of around 100,000 asylum cases and keep pace with tens of thousands of new applications each year. The stated goal was efficiency. But using such powerful, unpredictable technology to help decide whether someone is safe to return home has provoked alarm. Lawmakers, researchers, lawyers and judges have all raised questions. Now the Independent Chief Inspector of Borders and Immigration, the UK’s official watchdog for border and immigration matters, is considering an investigation. At the heart of the concern is a simple, terrifying problem: no one can guarantee that the AI is telling the truth, and the people whose lives depend on it have no way to know if it has lied.
The potential investigation has been building for months. The Independent Chief Inspector of Borders and Immigration, known as ICIBI, has contacted the digital rights nonprofit Open Rights Group (ORG) and asked it to share information about the tools as the watchdog prepares its inspection priorities for the next year. Sara Alsherif, a campaigner at ORG, confirmed that the request came after the group completed a six-month investigation into the Home Office’s AI systems. That investigation was painstaking. ORG sent eleven parliamentary questions, wrote two ministerial letters, and submitted ten public records requests in an effort to obtain basic information. The Home Office refused to disclose most of what it was asked, citing confidentiality and claiming that revealing details would undermine its operations. The secrecy has made independent verification almost impossible. The ICIBI says it cannot confirm what topics are under active consideration before its annual inspection priorities are published. But it has acknowledged the public concern and pointed to a previous finding that one of the tools’ “effectiveness and reliability were varied.” Those are not the words of an agency that trusts the technology completely. The Home Office, asked to respond, has insisted that the AI tools are simply assistants. Asylum decisions are made by trained caseworkers, a spokesperson said, and the tools are used only to help navigate interview records and policy information. They cannot be used in isolation and do not make decisions on claims. The department also says it regularly monitors the tools’ performance and has already committed to publishing updated evaluations. Yet the gap between the official line and the concerns raised by researchers is enormous, and the public is still waiting for the evidence that would settle it.
To understand why the Home Office turned to AI, you have to understand the state of the asylum system it is managing. For years, applicants have waited in agonizing limbo, unable to work, study, or plan, while their cases slowly moved through an overloaded bureaucracy. When the Labour government took office, it inherited a backlog of around 100,000 people waiting for an initial decision. Every year, tens of thousands more arrive and apply. Caseworkers are expected to process enormous volumes of information in a fair and careful way, while also dealing with the emotional weight of stories of persecution. The pressure is immense, and mistakes are inevitable. The AI pilot, launched in late 2024, was designed to relieve some of that strain. The Asylum Case Summarisation tool takes the transcript of the interview in which an asylum seeker explains why they fled, and compresses it into a shorter summary. The Asylum Policy Search tool acts as a research assistant: it searches official country policy documents and produces a summary, helping the caseworker decide whether an applicant would face genuine risk if returned home. In the pilot, using both tools together was estimated to save about an hour per case. That hour matters when you have tens of thousands of files waiting. The Home Office has pointed to progress: the number of people waiting for an initial decision was halved in the year through June. But that progress has a troubling shadow. Appeals against asylum decisions have ballooned. A decision made faster is not necessarily a decision made better, and the human cost of a wrong decision is not reflected in any performance dashboard.
At the core of the problem is the nature of the technology itself. Generative AI models like GPT-4 are not fact-checkers. They are statistical engines that predict words based on patterns learned from vast amounts of data. This makes them extraordinarily flexible—and extraordinarily prone to error. They can produce clear, confident, professional-sounding text that is entirely wrong. In the world of asylum, these errors are not annoying glitches; they are potential life-or-death mistakes. The Open Rights Group’s investigation found that the percentage of ACS-generated summaries deemed not fit for use was significantly higher than the 9 percent error rate reported in the official pilot evaluation. A summary can be deemed unfit because it contains a factual error, misrepresents what the asylum seeker said, or omits an important detail. When a caseworker relies on that summary, the original testimony may as well be invisible. The Asylum Policy Search tool, meanwhile, appears to have no quality control at all. ORG found no evidence that anyone checks whether the policy summaries it produces are accurate or complete. This is a massive gap in a system that is supposed to be careful. The dangers are not merely theoretical. In July, a British judge accused the Home Office of relying on an “AI hallucinated” document to refuse an asylum claim from a Moroccan woman and her child. The government claimed Morocco was safe for her to return to, but the document supposedly supporting that conclusion did not exist. The machine had invented it. The judge’s words should have been a wake-up call. Instead, the same tools continue to be used, quietly, behind closed doors.
Beyond individual errors lies a deeper problem of accountability. Asylum law exists to protect people who are in danger. It can only work if decisions are made fairly, based on reliable evidence, and open to challenge. But when an AI tool influences a decision, and the applicant cannot see what the tool produced, fairness becomes impossible. The Home Office has withheld information about the systems, resisted records requests, and published only limited evaluations. Lawyers representing asylum seekers have no way to know whether the summary of their client’s interview was accurate or whether the country policy summary that shaped the decision was based on invented material. They cannot mount an effective appeal because they are fighting an invisible adversary. This violates the basic idea of due process. It also puts an intolerable burden on applicants. They are asked to tell their stories, to prove their fear, to trust a system while that system uses hidden software to interpret and even distort their words. The power imbalance is enormous. An independent investigation by the ICIBI could change this dynamic. It would allow experts to examine the tools, test their outputs, and compare what the AI produces with what actually happened in individual cases. It could reveal whether the government’s claims about accuracy are true or whether the error rate is higher than officials admit. Most importantly, it would create a precedent: in a democratic society, the use of powerful technology in decisions that affect people’s fundamental rights cannot remain a secret. The state has a duty to explain itself, especially when it is deciding whether someone lives in safety or is returned to danger.
This is not simply a story about technology or bureaucracy. It is a story about human beings and the way they are treated when they are most vulnerable. The woman from Morocco whose child was refused protection because a machine invented a document is not a statistic. She is a person who trusted a system that failed her. The asylum seeker sitting in an interview room, reliving torture and loss, has no idea that a computer program may later decide which of their words matter. The caseworker, exhausted by a crushing caseload, may not know that the summary in front of them is built on guesswork. These are the faces behind the debate. The Home Office is under enormous pressure, and clearing a backlog of over 100,000 cases is a real challenge. But the answer cannot be to sacrifice justice for speed. If the tools actually help trained caseworkers make better decisions, then the government should be eager to show how they work and prove their reliability. Instead, it has chosen secrecy. That choice speaks louder than any press release. The Independent Chief Inspector of Borders and Immigration now has an opportunity to open the black box. By launching a full investigation, the watchdog can demand transparency, expose weaknesses, and insist that any use of AI in asylum decisions be safe, accurate, and accountable. The technology will not go away. It will only become more powerful. The question is whether it will be used in a way that respects the people it touches. Asylum seekers deserve careful, compassionate, truthful decisions. They deserve a system that sees them as individuals, not as cases to be processed. The investigation being considered now is a step toward that vision—if the government is willing to let the truth come out.
