UK Home Office asylum AI tools may face inspection, as data regulator already probes their use

On October 27, WIRED reported that two OpenAI-based tools used by the UK Home Office to process asylum claims may be added to the borders watchdog’s inspection priorities for next year. The Home Office is the government department responsible for UK immigration and asylum decisions; the two tools are Asylum Case Summarisation (ACS), which summarises asylum interview transcripts, and Asylum Policy Search (APS), which retrieves country-policy documents. The story lands on a hard deadline: OpenAI issued a deprecation notice for the GPT-4 family on October 14 and will retire the 2023-era model in January 2027 — and per WIRED, the government tools still run on it.

Key points

• WIRED reports the Independent Chief Inspector of Borders and Immigration (ICIBI) asked the Open Rights Group (ORG) for material while drafting next year’s inspection plan — but its exact words were that it “cannot confirm what topics are under active consideration.” No formal inspection has been ordered.
• A second line is already live: the Information Commissioner’s Office (ICO) has confirmed it is investigating the Home Office’s refusal to explain how staff use the two tools.
• ORG’s June report recorded that 9% of ACS summaries were unusable in the pilot evaluation and 23% of caseworkers lacked full confidence in the outputs; neither tool appears in the government’s Algorithmic Transparency Recording Standard (ATRS) repository.

Background

The tools were born out of backlog. When the pilot began in late 2024, around 100,000 asylum claims awaited an initial decision, and the Home Office’s own evaluation put the payoff at roughly an hour saved per case. The pilots ran and were evaluated between May and December 2024; the government published that evaluation on April 29, 2025. Per ORG’s report, APS then rolled out fully in July 2025, with ACS following in April 2026.

Scepticism built alongside the rollout. On March 16, ORG published a legal opinion by Robin Allen KC and Dee Masters of Cloisters Chambers, together with Joshua Jackson of Doughty Street Chambers: procedural fairness requires telling applicants that AI took part in processing their case, and not telling them is “likely to be unlawful.” The opinion also warned of a “significant risk that decisions based on those summaries will be based upon and vitiated by material errors of fact.” On June 18, ORG and SOAS’s ICOP project published the report naming the transparency and accountability failures. The Register covered the same investigation on October 13 under the headline “one in five accuracy checks.” WIRED’s October 27 story adds the regulators’ stance.

The facts

There are three sets of accuracy numbers, from different sources at different stages, and they don’t average together.

The pilot evaluation (gov.uk, published April 2025) found 9% of ACS summaries were flagged as unusable in quality checks; ORG’s report adds that 23% of caseworkers said they lacked full confidence in the outputs. After full rollout (The Register, citing ORG’s investigation into post-April-2026 expert spot checks), 45 of 203 AI-generated interview summaries failed initial expert checks — about 22% — with another 19 flagged for review, and 30 of 56 second-stage reviews failing the required standard. ORG says real-world error rates run significantly above the published 9% but gives no overall figure. Coverage is another gap: ORG estimates experts checked only about 0.3% of summaries between April and July, against a projected monthly caseload of 17,515.

Where regulators and lawyers stand:

Item	Status	Source
ICIBI inspection	Not opened; “aware of concerns,” cannot confirm topics under consideration	WIRED, Oct 27
ICO probe	Open: the Home Office’s refusal to disclose how staff use and are trained on the tools	WIRED, Oct 27
Legal opinion	Not telling applicants “likely to be unlawful” (2026-03-16)	ORG press release
ATRS listing	Neither tool registered in the transparency repository	ORG report
Suspension call	ORG demands immediate suspension and an independent audit	ORG report

On the model version, sources conflict and cannot be merged: ORG’s report says both tools run on GPT-4o (APS from July 2025, ACS from April 2026), and The Register likewise describes GPT-4o via Microsoft’s Azure AI Foundry; WIRED says the tools are built on OpenAI’s 2023 GPT-4 running on Microsoft Foundry. Which is right should come out in official documentation or the ICO’s findings. What is certain is the retirement timeline itself: OpenAI’s October 14 notice retires the GPT-4 family in January 2027.

The Home Office response stays on two lines: AI is “just one of the tools at our disposal to protect our borders, and does not replace or automate the role of human decision-makers,” and asylum decisions are made by trained caseworkers with AI used only to assist in navigating records and policy information — “they cannot be used in isolation and do not make decisions on claims.” The department also says it has committed to publishing updated evaluations of the rollout.

What others say

Civil society is aligned and escalating. Sara Alsherif, who leads ORG’s migrant digital justice programme, told WIRED that migrant organisations and lawyers largely don’t know the tools exist: “we need to let them know how it works, so they can challenge it.” ORG’s formal demands: suspend both tools immediately, publish the data protection and equality impact assessments, and register the tools in ATRS. The Public Law Project adds context: ACS and APS are two of eleven high-risk AI systems in use across government, yet absent from the public repository.

The regulators are far more careful. ICIBI told WIRED it is “aware of concerns about the use of AI tools” and cited earlier inspection findings that one tool’s “effectiveness and reliability were varied” — while saying it “cannot confirm what topics are under active consideration” before the annual priorities are published. That hedge is exactly where the “may face” in this story lives. The two outlets frame the story differently, too: The Register leads with error rates and check coverage, WIRED with the regulatory trail and the model retirement. Their overlap is the transparency gap.

Our take

The general lesson here is about where “AI-assisted decisions” end. The Home Office insists humans decide; ORG and the legal opinion point at what happens before the decision — applicants are never told their interview was summarised by a model, so they can neither know nor correct a wrong summary. If errors enter upstream of the decision and stay undisclosed, “a human signs off” is not procedural fairness. For teams wiring large models into government or enterprise workflows, the concrete lesson is that transparency registers like ATRS are not paperwork: when the regulators and the press finally look, the absence is the story. Compared with California’s No Robo Bosses Act, what’s missing on the UK side is the same thing that law mandates — disclosure and human review.

The model question is a real engineering clock. If WIRED’s GPT-4 reading is right, the January 2027 retirement means the government must swap the foundation model within months — and any model swap changes the summaries, which means re-validating whether the error rate is 9% or something worse. The honest caveats: the “may face investigation” line rests on WIRED alone; ICIBI has not opened anything; ORG gives no overall error-rate number, so none is invented here; and the model version has two conflicting accounts. All of that is on the record above, with sources, for readers to check.

What to watch

• When ICIBI publishes its annual inspection priorities: whether asylum AI appears, turning “may” into a formal inspection.
• The ICO investigation’s progress, and whether — and when — the Home Office publishes its promised updated rollout evaluations.
• January 2027, when the GPT-4 family retires: the government tools’ migration and any re-testing of accuracy numbers.