UK watchdog may probe secret OpenAI tools used in asylum decisions

The UK’s border-inspection watchdog is considering a review of the Home Office’s secret use of OpenAI-based tools in asylum processing, after campaigners said the department has not properly disclosed how the systems are used, trained or checked. The possible inspection matters because the tools are already embedded in a high-stakes part of government decision-making where errors can affect whether people remain in the UK.

What is being examined?

According to WIRED, the Independent Chief Inspector of Borders and Immigration, known as ICIBI, is weighing an inspection into the Home Office’s use of artificial intelligence in asylum work. The issue centres on two tools, Asylum Case Summarisation and Asylum Policy Search, which were piloted in late 2024 as the government looked for ways to reduce a large backlog of asylum cases.

The Home Office has not publicly disclosed the full details of the tools, WIRED reported, even though the department says they are only there to assist trained caseworkers rather than make standalone decisions. That gap between official assurances and public disclosure is what has put the system under scrutiny.

Why it matters

At stake is not just whether AI can speed up administrative work, but whether the use of that technology is visible, auditable and fair in a process that can shape a person’s legal status and future. Asylum applicants are not told when the tools are used in their case, according to WIRED, which means they may have little practical way to know whether a summary or policy search influenced a decision they want to challenge.

The concern is amplified by the nature of asylum decision-making itself. It depends on careful reading, legal judgment and the accurate handling of personal evidence. If a system produces a poor summary, misses context or surfaces the wrong policy, the result can be a flawed decision that is difficult to unwind later.

Campaigners also argue that the issue is larger than one department. The Public Law Project has said 11 high-risk AI systems across government are missing from the public repository, underscoring a wider transparency problem around how public bodies are deploying automated tools.

How does the Home Office say the system works?

The Home Office says AI does not decide cases on its own. A spokesperson said:

Asylum decisions are made by trained caseworkers, with AI tools used only to assist…

That framing is important, but it does not resolve the transparency question. A tool can shape a decision without formally making it. In asylum cases, a system that condenses evidence, flags key points or searches policy documents can still affect what a caseworker sees first, what gets overlooked and how quickly a file moves.

According to documents obtained by the Open Rights Group, the two tools were built on OpenAI’s GPT-4 and Microsoft Foundry. ORG says it pieced together the picture through 11 parliamentary questions, two ministerial letters and 10 records requests, which suggests the public record has been assembled slowly and only in fragments.

Background: how we got here

The Home Office began piloting ACS and APS in late 2024, at a time when the asylum system was under intense pressure to clear a backlog of roughly 100,000 cases. The government later said the number of people waiting for an initial asylum decision fell by half in the year through June, a sign it views throughput as an urgent policy goal.

But speed has come with new questions about reliability. WIRED reported that appeals against asylum decisions have risen sharply, and campaigners say that should prompt a closer look at whether faster processing is translating into better decisions or simply faster ones.

The public debate has also been shaped by earlier concerns over AI accuracy in immigration work. A UK judge previously said the Home Office cited an AI-hallucinated document in a Moroccan asylum refusal, a reminder that even isolated errors can have serious consequences when they enter a legal process.

What are ACS and APS?

ACS and APS are the Home Office’s internal AI-assisted tools for asylum processing. According to the documents cited by ORG, ACS is used for case summarisation and APS for policy search, with the systems built on OpenAI and Microsoft infrastructure.

Wired reported that the government estimates the combined use of the tools can save about an hour per case. That kind of time saving explains the attraction: a system that trims manual reading or policy lookup can, in theory, help caseworkers handle more files. The trade-off is that the public is being asked to accept a process that remains largely hidden.

What investigators and campaigners are saying

ICIBI has not confirmed any active topic ahead of the publication of its annual inspection priorities, but it has acknowledged that it is aware of concerns about AI use. WIRED reported that the inspectorate pointed to last year’s mixed finding on effectiveness and reliability, signalling that AI-enabled systems are already on its radar even if this specific review is not yet official.

Open Rights Group argues that the lack of disclosure undermines the ability of asylum seekers to understand and challenge decisions. Campaigner Sara Alsherif put it bluntly:

So we need to let them know about this, let them know how it works, so they can challenge it.

Labour backbench MP Bell Ribeiro-Addy also sharpened the political stakes, saying:

We’re dealing with people’s lives.

Those remarks capture the central tension in the story. The Home Office is trying to move faster. Critics say speed cannot come at the expense of due process, especially when the people affected may not even know an AI tool was involved.

Transparency gaps and oversight questions

The case against the Home Office is not only about the tools themselves, but about what has not been published. WIRED reported that the department has not publicly disclosed the full details of ACS and APS. It also said the tools are absent from the public repository required under the government’s AI Playbook for algorithmic decision tools.

That omission matters because transparency frameworks only work if the public can see what systems exist, what they do and where they are used. If a high-risk tool is operating inside a sensitive workflow but is not visible in the official register, that weakens external oversight and makes it harder for journalists, MPs and civil society groups to assess risk.

The Information Commissioner’s Office is already investigating the Home Office’s refusal to explain staff use and training on the tools, WIRED reported. That adds another layer of scrutiny: even if ICIBI does not move ahead immediately, data protection and transparency regulators are already asking questions.

Why might the inspectorate act now?

Because asylum processing has become one of the most visible test cases for government AI in the UK. ICIBI is responsible for inspecting Home Office border and immigration operations, including the management of asylum processes, so the tools sit squarely within its remit.

The inspectorate may also be responding to a broader shift in public expectations. Government departments increasingly use algorithmic tools to speed administrative work, but the threshold for disclosure is higher when those tools touch rights, eligibility and legal status. That is especially true when the department says the systems are only assistive, which makes independent scrutiny of how they are actually used even more important.

By the numbers

Comparison of the main claims

What the Home Office says versus what campaigners say

The divide is straightforward. The Home Office says trained staff make the decisions and AI only helps them work faster. Campaigners say the department has not been transparent enough about how the systems operate, what safeguards exist and whether people affected can find out when AI is involved.

WIRED reported that ORG also says APS has no quality control for generated summaries, while ACS has a summary rejection rate above the published 9% pilot error rate. Those are significant allegations because they go to the heart of whether the tools are merely administrative aids or unacceptably unreliable inputs into legal decisions.

What happens next?

The immediate question is whether ICIBI will formally launch the inspection into the Home Office’s AI use. WIRED said the watchdog is considering it, but has not yet confirmed the topic because annual priorities have not been published. That means the timing remains uncertain, even if the issue is already under active review.

Another thing to watch is whether the Home Office publishes the updated evaluations it says are coming. If those reviews are detailed enough, they could answer some of the questions around accuracy, oversight and training. If they are vague, pressure for a formal inspection will probably increase.

There is also the Information Commissioner’s investigation, which could force the department to explain more about staff use and training. That inquiry may end up being important not only for this case but for any future attempt to use AI in immigration work without clear public disclosure.

Finally, there is the broader policy question of whether other departments are using similar undisclosed tools. If the asylum system is the first visible flashpoint, it may not be the last. The real test is whether government can use AI to improve administration without eroding the transparency and challenge rights that make administrative decisions legitimate in the first place.

Background context: why asylum is such a sensitive AI test

Asylum processing is not a routine back-office task. It involves facts, credibility judgments, legal standards and life-changing consequences. That combination makes it one of the hardest places to introduce automated support tools without triggering concerns about bias, error and accountability.

The UK already has a policy environment that expects transparency for high-risk decision tools, and that is why the absence of ACS and APS from the public AI repository is drawing attention. When a public authority uses software to help decide cases involving protection claims, the burden should be on the state to explain how the system is controlled and how errors are corrected.

That pressure is only likely to grow if, as WIRED reported, appeals are rising sharply even while the initial-decision queue has fallen. Faster output is not the same as better adjudication. If the use of AI is helping the Home Office process more files, the public will still want to know whether the results are accurate, whether the underlying summaries are fair and whether people can discover when a machine has touched their case.

For now, the watchdog review remains only a possibility. But the paper trail already points to a familiar modern governance problem: technology introduced to speed up the state has moved faster than the systems meant to keep it visible.