A six-month investigation by digital rights organization Open Rights Group has found that the UK Home Office's AI asylum tools are generating errors in more than one in five spot-checked outputs — and that quality testing on the second tool has stopped entirely. Both tools run on OpenAI's GPT-4o through Microsoft's Azure AI Foundry, are embedded in every asylum case the UK processes, and have never been disclosed to the applicants whose testimony they summarize. The full investigation, "Automated Discrimination: Unpacking the Digital Hostile Environment," was published today by ORG after a six-month inquiry combining freedom of information requests, parliamentary questions, and expert legal analysis.

The findings land at a moment when the broader question of how far governments can responsibly delegate consequential decisions to generative AI — and at what error tolerance — is being actively contested in courts, parliamentary committees, and academic literature worldwide. The UK Home Office's situation offers one of the most detailed public records of what that delegation looks like in practice: error rates more than doubling from pilot to production, oversight sampling less than a third of one percent of outputs, no independent audit planned, and no notification to the people most affected.

Two Tools, One Architecture, Expanding Errors

At the center of the controversy are two generative AI systems the Home Office has embedded in every asylum case it processes: the Asylum Case Summarisation tool, known as ACS, and the Asylum Policy Search tool, known as APS. Both are built on OpenAI's GPT-4o large language model, delivered through Microsoft's Azure AI Foundry enterprise cloud service.

APS was rolled out to all asylum decision-makers in July 2025. It provides a chat-style interface through which caseworkers can query the government's country policy information — background documents used as evidence in asylum hearings that describe conditions in applicants' countries of origin. ACS followed in April 2026, automating the summarization of asylum interview transcripts, which can run to 50 pages or more, before decision-makers review them.

The technical distinction that drives ORG's concern is foundational: both tools are generative, not retrieval-based. A retrieval system surfaces what is already in the source document; a generative system rewrites, synthesizes, and may filter or omit facts from that document depending on how the model's outputs diverge from the underlying record. For asylum determinations — which are built on detailed accounts of individual testimony — that distinction matters in ways that a spell-checker or formatting tool would not.

Error Rates More Than Doubled from Pilot to Production

The Home Office's own pilot evaluation of ACS, published in April 2025, found that 9 percent of summaries were inaccurate or omitted information significant enough to require removal from the trial. That figure — already a cause for concern in a high-stakes legal context — appears to have worsened substantially once the tool was deployed at scale.

Documents obtained by ORG through freedom of information requests show that Home Office subject matter experts subsequently conducted spot-check sampling of 203 ACS summaries generated between April and July 2026. Summaries were classified as failed if they contained what the Home Office's own criteria described as "multiple factual errors or misrepresentations" or "significant inaccuracies, contradictions, or misstatements." In the first stage of that review, 45 of the 203 summaries — roughly 22 percent — failed outright, with a further 19 flagged for additional scrutiny. A second-stage review found 15 percent of sample summaries failing the quality bar.

The gap between the 9 percent pilot figure and the 15-22 percent range observed in production is not unusual for generative AI systems: model behavior can shift when the distribution of inputs in live deployments diverges from pilot conditions. What is unusual, ORG argues, is that the Home Office deployed the tool at full scale while aware of the pilot error rate, without establishing quality-control systems capable of detecting that apparent deterioration.

A further limitation compounds the concern. Under the Home Office's failure criteria, a summary that omits material facts about a case can still pass quality review if the content it does include is accurate. The relevance of what the model chose to include is not assessed as a grounds for failure — meaning a summary that presents an accurate but incomplete picture of an applicant's testimony could be rated acceptable even if the omitted information was the most significant element of the case.

How Do You Audit an AI You Test in 0.3 Percent of Cases?

The 203 summaries reviewed by subject matter experts represent approximately 0.3 percent of the roughly 70,000 asylum cases processed by the Home Office between April and July 2026, the months after ACS was rolled out to all caseworkers. The agency's own data protection impact assessment for ACS reports that 17,515 cases are processed on average each month, putting the scale of unaudited AI-generated output into sharp relief.

At a 22 percent first-stage failure rate, the implication is straightforward: approximately 3,500 summaries per month that would fail quality review are never checked. The statistical inadequacy of the sampling regime is not merely a matter of resource constraints — it is a structural consequence of the sampling size. With n=203 and a 22 percent observed failure rate, the 95 percent confidence interval for the true underlying error rate spans roughly 16.6 percent to 27.4 percent. The regime cannot distinguish an 18 percent error rate from a 26 percent error rate with any meaningful confidence.

ORG further reports that the outcomes of the sampling exercise have been "mainly discussed verbally" within an internal working group, rather than documented in published assessments. An interim spot-testing report has been provided to an internal board but has not been made public. The Home Office has confirmed to parliamentarians that it has no plans for an independent audit of the ACS system.

Why GPT-4o Is Structurally Prone to Errors in This Context

The error rates ORG documents are not primarily an implementation problem — they reflect an architectural property of the model class both ACS and APS employ. Large language models such as GPT-4o are generative systems trained to produce fluent, coherent text by predicting likely continuations of input sequences. They do not retrieve facts from a source document and transcribe them; they synthesize new text that reflects the statistical patterns in their training data and in the immediate prompt context.

This means GPT-4o can produce text that is both fluent and factually incorrect — not because of a bug, but because fluency and factual accuracy are different properties that training does not guarantee will align. In legal and research literature, this phenomenon is called hallucination: the model presents fabricated or distorted content with the same confidence as accurate content. A global legal hallucination tracking database maintained by researchers at HEC Paris and Sciences Po has documented more than 1,300 such cases across multiple jurisdictions. In July 2026, a UK judge specifically cited AI hallucination as a likely cause of an asylum refusal based on nonexistent policy — a country policy note that did not exist.

Deploying GPT-4o in a production environment does not eliminate this tendency; it manages it through prompt engineering, testing, and quality controls. The ORG investigation's core finding is that the Home Office's quality controls — sampling 0.3 percent of outputs, with no independent audit and no ongoing expert testing of APS — are structurally insufficient to manage the architectural risk of the model class it has chosen.

APS: Testing Has Stopped Entirely

The situation for the Asylum Policy Search tool is, in ORG's assessment, worse than for ACS — because no equivalent ongoing quality-control process has ever been established. APS operates as a generative interface over country policy documentation, functioning as an AI assistant that caseworkers query to find relevant information about conditions in a given country. That information forms a significant evidential basis for asylum decisions.

Anna Turley, the Minister of State for Border Security and Asylum, confirmed in August 2026 that expert testing of the APS tool had ceased entirely. ORG says it found no evidence that any systematic quality checking of APS outputs has been conducted since full rollout. The Home Office has not provided ORG or parliamentarians with the prompts used to instruct the model, the system's internal evaluation criteria, or details of how caseworkers are trained to recognize and handle errors in its outputs — information the department declined to disclose in response to freedom of information requests.

The cessation of APS testing is significant because the tool's error rate in production is entirely unknown. Unlike ACS, where the 22 percent figure at least quantifies the problem, APS has produced no equivalent data since it went live in July 2025. The tool has been in daily use for more than 15 months without any confirmed ongoing quality mechanism.

A Removed Layer of Expert Scrutiny

ORG's investigation draws attention to an additional consequence of the ACS rollout that has received less public attention than the raw error rates. According to the tool's data protection impact assessment, ACS was developed to automate work previously performed by technical specialists — staff who read and summarized full case files before passing them to decision-makers. The implication is that those specialists no longer perform that full-file review; the AI summary has replaced it.

If accurate, that substitution removes a layer of human expert scrutiny that previously existed independently of the decision-maker. The Home Office has maintained that caseworkers who use ACS must also read every underlying document. But ORG says this claim is difficult to reconcile with the productivity gains the department cites as justification for the tool — savings of 23 minutes per case in a system where reading a 50-page transcript in full would take considerably longer than the time reportedly saved.

Monthly initial asylum decisions also fell roughly 42 percent in April-July 2026 compared with the same period in 2025, while mean interview-to-decision time rose 29 percent. The efficiency gains the Home Office projects do not appear to have materialized in aggregate output data — suggesting that the deployment's effects on decision-making time are more complex than the per-case savings figure implies.

Neither Tool Appears on the Government's Own AI Registry

Two structural transparency failures run through ORG's findings. First, asylum applicants are not told when AI tools are used to summarize their testimony or to search policy documentation relevant to their case. ORG's March 2026 legal opinion, authored by Robin Allen KC, Dee Masters of Cloisters, and Joshua Jackson of Doughty Street Chambers, concluded that this failure is likely unlawful under existing UK data protection law. The Home Office has said its privacy notice explains that parts of processing may involve automation, but ORG and the legal opinion argue this generic disclosure is insufficient to satisfy data rights obligations for a processing activity of this consequence.

Second, neither ACS nor APS has been listed on the Algorithmic Transparency Recording Standard Hub — the government's own public registry of AI tools in use by public bodies. Caroline Selman, a senior research fellow at the Public Law Project, has documented that ACS and APS are among 11 high-risk AI systems in use across government that are absent from the public repository. The absence means that the public, parliamentarians, and legal representatives acting for asylum applicants have had no official notice that the tools exist, let alone what their error rates are.

"Deciding whether someone can or cannot seek refuge in the UK is one of the most serious and life-changing decisions that the state can make about an individual," said Sara Alsherif, ORG's migrant digital justice program manager. "The processes that lead to that decision must be transparent, fair and accountable. But the Home Office continues to test AI on people seeking asylum in the UK without even telling them."

Regulators Begin to Circle

The report's publication has prompted two oversight bodies to act. The Independent Chief Inspector of Borders and Immigration told ORG that it is seeking additional information from the organization as it develops inspection priorities for 2027 — a signal that a formal inspection of the Home Office's use of AI tools in asylum casework is under consideration. An ICIBI inspection would carry significant weight: its reports are presented to Parliament, and the Home Office is required to respond publicly to its recommendations.

Separately, the Information Commissioner's Office is investigating the Home Office's conduct in declining to disclose staff training details for ACS and APS. That investigation concerns whether the department's handling of ORG's freedom of information requests conformed to its obligations under UK data rights law. If the ICO finds against the Home Office, it could require disclosure and potentially trigger a broader audit of the department's AI governance.

The dual scrutiny reflects a broader pattern ORG documents throughout its investigation: the Home Office has consistently invoked national security exemptions to resist transparency requests about its use of AI across immigration functions. ORG's six-month inquiry depended on a combination of ten public records requests, 11 parliamentary questions, and two ministerial letters to assemble the picture that its report now presents.

A Wider AI Infrastructure Built Without Disclosure

The report's scope extends beyond ACS and APS. "Automated Discrimination" maps what ORG describes as a rapidly expanding catalogue of AI deployments across the UK's immigration system, embedded with what the report calls "near total secrecy." These include AI-powered live facial recognition and drone surveillance at the UK's borders; a system for estimating the ages of asylum seekers without a disclosed technical basis; AI-assisted searching of the mobile phones of new arrivals; and the Identify and Prioritise Immigration Cases tool, or IPIC, an algorithmic case-prioritization system whose criteria have not been publicly disclosed.

Report author Fabio Chiusi, a journalist and researcher whose work focuses on AI and society, frames the accumulation of these tools in explicitly political terms: the report argues that the framing of migration as a permanent emergency to be managed by algorithmic systems normalizes the so-called hostile environment first articulated by then-Home Secretary Theresa May in 2012. Labour MP Bell Ribeiro-Addy, who assisted ORG's investigation, said the report represents an opportunity "to break with the Home Office's toxic record of failure." SNP MP Pete Wishart separately criticized the department's failure to notify asylum applicants when AI is used in their cases.

The Home Office Responds

A Home Office spokesperson, responding to the report's findings, maintained that human decision-makers remain in control. "Asylum decisions are made by trained caseworkers, with AI tools used only to assist in navigating interview records and policy information. 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 and have already committed to publishing updated evaluations on their rollout."

The department did not directly address ORG's figures on error rates, the cessation of APS expert testing, the absence of both tools from the ATRS registry, or the question of whether technical specialists continue to conduct full-file reviews independently of the AI summaries.

What ORG Is Demanding

ORG is calling on the government to suspend both ACS and APS pending an independent audit and to implement an outright ban on the use of generative AI tools in asylum applications — a domain it argues carries risks to fundamental rights too severe to permit the current level of error tolerance. Its broader policy demands include stronger freedom of information legislation, a narrower statutory definition of national security that cannot be routinely invoked to block disclosure of AI governance details, and a mandatory public register of government AI systems with sufficient technical detail to enable independent assessment.

The tools at the center of the controversy are not exotic frontier models. They run on a commercially available large language model through a standard enterprise cloud service. The technical stack is unremarkable. What is remarkable is that tools generating errors in somewhere between 15 and 22 percent of sampled outputs — with spot-check coverage of less than a third of one percent of production volume, no independent audit planned, and no notification to affected individuals — have been rolled out into every asylum case the UK government processes.

As AI systems become faster, cheaper, and more capable, the governance question is whether public-sector institutions can build accountability frameworks fast enough to keep pace. The ORG investigation suggests that, in at least this corner of the UK government, they have not yet come close.

Frequently Asked Questions

How does AI affect asylum decisions in the UK?

The Home Office uses two generative AI tools — the Asylum Case Summarisation tool (ACS) and the Asylum Policy Search tool (APS) — in every asylum case it processes. ACS generates summaries of interview transcripts that can run to 50 pages or more; APS answers caseworker queries about country policy documentation used as evidence in hearings. Both tools are built on OpenAI's GPT-4o and run through Microsoft's Azure. The Home Office maintains that caseworkers make all decisions and must read underlying documents, but the tools' error rates and the removal of the independent technical specialist review raise questions about how that oversight functions in practice.

What is the Asylum Case Summarisation tool, and how accurate is it?

ACS is a generative AI tool that produces summaries of asylum interview transcripts before they reach decision-makers. During a pilot evaluation in 2025, 9 percent of summaries were found to be inaccurate or to omit significant information. In production spot-checks covering April-July 2026, that figure rose to 22 percent at first-stage review and 15 percent after a second pass — a range representing the Home Office's own quality criteria. The tools do not assess whether summaries omit relevant facts, only whether facts they include are accurate, meaning a selectively incomplete summary can still pass quality review.

Is the Home Office AI asylum tool legal?

A March 2026 legal opinion commissioned by ORG from Robin Allen KC, Dee Masters, and Joshua Jackson concluded that the Home Office's use of ACS and APS is "likely unlawful" under existing UK data protection law, primarily because asylum applicants are not informed when AI is used to process their cases and therefore cannot exercise their data rights. The Information Commissioner's Office is separately investigating the Home Office's refusal to disclose staff training details for both tools. No court has yet ruled on the question, and the Home Office disputes the legal analysis.

Can asylum seekers challenge decisions where AI was used?

Asylum seekers in the UK are not notified when ACS or APS is used in their case. Without that disclosure, they and their legal representatives cannot identify whether an AI-generated summary or policy search contributed to an adverse decision, or challenge its accuracy. ORG's legal opinion argues this notification gap is itself the core unlawfulness: data protection rights cannot be exercised against processing that is not disclosed. Legal representatives who suspect AI involvement can raise it in appeal proceedings, but without formal disclosure they are dependent on what emerges through freedom of information requests or parliamentary questions.