AI in UK Asylum Decisions: Accuracy and Oversight

What AI tools is the Home Office using?

The Home Office says it uses two tools for asylum casework: Asylum Case Summarisation (ACS) and Asylum Policy Search (APS). According to documents obtained by the digital rights nonprofit Open Rights Group (ORG), both were built using OpenAI's GPT-4 model and Microsoft Foundry. The department began piloting them in late 2024, aiming to reduce a backlog of asylum claims while also processing new applications.

The tools do different things. ACS summarizes interview transcripts where applicants explain what happened to them and why they fear returning. APS helps staff search for and summarize country policy information that may be relevant to whether someone faces a real risk of harm.

That distinction matters. One tool is working from a person's account; the other is designed to pull in context about the country in question. Neither is the same as making the decision itself - but both can influence what a caseworker notices and how they interpret it. A summary can steer attention toward certain details. A search result can affect how conditions in a country are portrayed.

The Home Office says asylum decisions remain the responsibility of trained caseworkers, and that the tools are not meant to be used on their own to decide claims. It also says it monitors tool performance through existing quality-assurance processes and plans to publish updated evaluations. Those points don't, by themselves, settle whether the tools perform reliably across real cases.

The efficiency argument is straightforward. Government estimates from the pilot suggested that using both tools could save about an hour per case. In a high-volume system, that's not trivial. When the tools were introduced, the UK had around 100,000 asylum cases awaiting an initial decision, and the Home Office later reported a reduction in the number waiting. But saving time only helps if the work stays accurate - and if the time saved doesn't depend on a less careful review.

A clearer way to think about the system is as a chain, not a single AI verdict:

• An applicant provides evidence, including an interview account.
• ACS produces a summary that may guide a caseworker's review.
• APS helps staff search for relevant country information.
• A caseworker assesses the evidence and makes the decision.
• Quality controls should identify errors and provide a way to challenge them.

The risk can show up at any point in that chain. Calling the technology an "assistant" doesn't remove its influence - it just makes it more important to scrutinize the design of each step and the checks between them.

Why are accuracy and fairness concerns serious?

The concern isn't that every AI summary is wrong. It's that errors in a high-stakes process can have real consequences - and that the evidence available so far raises questions about how often those errors occur and whether they get caught.

ORG's investigation found that the proportion of ACS summaries judged not fit for use was higher than the 9 percent error rate reported in the pilot evaluation. ORG also said APS summaries lacked a quality-control process. The information available doesn't prove that every flagged problem changed an outcome. Still, a difference between a published pilot measure and later findings is a reason to look closely at how the evaluation was done, how the sample was chosen, and what checks actually happen during day-to-day use.

Summarization has a particular failure mode: it can sound coherent while omitting, reshaping, or misstating details that matter. In an asylum interview, chronology, a person's wording, uncertainty, and context may all count. If a summary smooths over a contradiction, leaves out an explanation, or presents something uncertain as if it were definite, a caseworker may start from a distorted version of the record. And when the distortion comes in the form of a polished paragraph, it can be harder to spot than an obviously incomplete document.

Country information brings a different set of risks. A search assistant can return a source that's outdated, irrelevant, or described incorrectly. In July, a UK judge accused the Home Office of relying on an AI-hallucinated document when refusing the asylum claim of a Moroccan woman and her child, on the grounds that Morocco would be safe for them. That reported case is a reminder that references generated or surfaced by AI still need to be verified against real sources. It doesn't, by itself, show how common this kind of mistake is - but it does show what it can look like.

Fairness is also practical, not just theoretical. ORG, the Public Law Project, and Labour MP Bell Ribeiro-Addy have raised concerns that large language models could introduce cultural bias in how language and context are represented. If someone's account is summarized in a way that strips away nuance - or if an output makes some details look less credible - a caseworker may have to work harder to recover the meaning from the original evidence.

That creates a distributional risk: not every applicant is likely to be affected in the same way. People who rely on interpreters, use unfamiliar terms, or describe experiences that aren't well represented in the data and assumptions behind a model may be more exposed to misrepresentation. So the relevant test isn't just whether "the average" output seems acceptable. It's whether error rates and performance are measured across different types of cases and different language situations the system actually encounters.

What is known about transparency and accountability?

The watchdog question is still open. The Independent Chief Inspector of Borders and Immigration (ICIBI) has asked ORG for information as it considers topics for future inspections. At the same time, the inspectorate hasn't confirmed that it will investigate the tools. It has said it's aware of concerns about AI use, and pointed to earlier findings that tool effectiveness and reliability varied.

That difference matters: a possible inspection is not the same thing as a finding that something went wrong. It can simply mean that there's enough uncertainty around the system to justify independent scrutiny. ORG's investigation involved parliamentary questions, ministerial letters, and public-records requests, reflecting how difficult it has been for outside groups to obtain details about how the tools work and how they're used.

There are also gaps in public disclosure. The UK government's AI Playbook says departments should be transparent about algorithmic tools used in decision-making and provide information through a public repository. ORG says ACS and APS are not listed there. The Public Law Project has also identified 11 high-risk AI systems it says are missing from that repository, raising a broader concern that what researchers know about may not reflect the full picture.

According to ORG campaigner Sara Alsherif, applicants aren't told when these tools are used in their cases. That limits what they can do to challenge whether a summary contains an error, or whether information from a search tool influenced the assessment. Lawyers at Cloisters, in an opinion commissioned by ORG, argued that blocking people from contesting incorrect AI summaries of their personal information may breach UK data-rights law. The Home Office's privacy notice says parts of the decision-making process may involve automation, but that's not the same as giving an applicant practical, case-specific information about how a tool affected their file.

Both ORG and the Public Law Project have sent their findings to the UK data regulator, which is investigating the Home Office's refusal to explain how staff use the tools and how staff were trained to use them. That process is separate from any possible ICIBI inspection. Different oversight bodies may focus on different questions: whether the department disclosed information, whether personal data rights are respected, and whether the system works reliably in practice.

The Home Office says human decision-makers remain responsible and that AI doesn't replace their role. That's true - but responsibility only matters if staff have the time, training, and access they need to verify outputs. If a caseworker can't view the original evidence alongside a summary, or if they don't know when a model generated a claim, "human sign-off" can end up being more procedural than meaningful.

Transparency doesn't mean publishing every technical detail or exposing sensitive personal data. It means making the system inspectable: naming what tools are used, explaining their role, documenting their limits, and giving affected people a way to raise concerns. Without those basics, applicants and the public can't judge whether safeguards match the stakes.

What would effective oversight look like?

A credible oversight system should test the whole workflow, not just whether a model can produce fluent text. For ACS, the key question is whether the summary preserves the material details of an interview. For APS, it's whether the system directs staff to relevant, verifiable country information and represents it accurately. In both cases, evaluation should focus on the caseworker's task - not on a generic score for model quality.

A practical evaluation would include controls like these:

• Compare outputs with source records. Check summaries against original interviews and cited material, paying close attention to omissions, changed meaning, and claims that aren't supported.
• Measure errors by type and context. Track factual inaccuracies, missing details, misleading framing, and unverified references separately. Then break results down by language, interpretation, and case characteristics that might affect performance.
• Test in real operational conditions. A pilot sample may not match later workloads, changing case types, or staff practices. The evaluations shouldn't stop after deployment begins - they should be repeated.
• Record how outputs are used. Keep an audit trail showing when a tool was used, what source material it processed, what it produced, and what the human reviewer checked or corrected.
• Set clear escalation rules. If an output can't be verified, includes a disputed detail, or points to a source that can't be confirmed, it shouldn't be treated as evidence until it's reviewed further.

Human review also has to be designed, not just assumed. A reviewer should be able to view the source material next to the AI output, tell the difference between generated text and verified evidence, and record what gets corrected. Training should cover known failure modes - including omissions and fabricated references - not only how to operate the software. Managers also need to make sure workload targets don't squeeze out the time needed for meaningful review.

People affected by the process need a way to find out whether AI-assisted tools were involved and to challenge relevant inaccuracies. That doesn't require everyone to get a technical description of the model. It does require a process that makes it possible to identify an error, understand how it entered the case record, and ask for it to be corrected before it weighs in on the decision. Legal advisers also need enough information to raise a concrete challenge.

Public reporting should make the evidence easy to follow. Publishing an error rate without explaining the sample, the definition of an error, the review method, and the limitations doesn't let readers judge reliability. And saying tools are monitored is less helpful than reporting what's monitored, how often, what thresholds trigger action, and what changes follow when problems are found.

There's a trade-off. More review and disclosure take time, and extra checks can reduce some of the efficiency gains. But the alternative isn't cost-free. An undetected error can lead to a bad decision, a challenge, additional administrative work, and a loss of confidence in the system. The comparison shouldn't be "AI speed versus no technology." It should be "assisted decisions with real safeguards versus decisions based on weak verification."

Finally, systems should be reassessed when their underlying model or configuration changes. ORG's reporting notes that OpenAI has discussed retiring GPT-4. A replacement model may behave differently, which means evaluations tied to one version can't automatically validate another. In practice, the department should treat model changes as a reason to test again, publish the results, and update staff guidance.

Key takeaways

• ACS summarizes asylum interview records, and APS helps staff search and summarize country policy information.
• The Home Office says trained caseworkers make decisions, but AI-generated material can still shape what caseworkers review and how they understand it.
• ORG reported a higher rate of ACS summaries judged unfit for use than the pilot's published 9 percent error rate, and said APS summaries lacked quality control.
• A potential ICIBI inspection is being considered, not confirmed. Separate scrutiny of data rights and disclosure is also underway.
• Strong safeguards require source-level verification, measurable error reporting, audit trails, staff training, and a practical way for applicants to challenge inaccuracies.