Dispute over the UK Home Office's use of generative AI to summarise asylum interviews

On 12 October, the UK digital rights group Open Rights Group (ORG) published a report pulling together six months of investigation into two generative AI tools used by the Home Office. The Asylum Case Summarisation (ACS) tool turns asylum interview transcripts into summaries for decision-makers, and the Asylum Policy Search (APS) tool lets decision-makers search country policy information through a chat-style interface. ORG's report gives OpenAI's GPT-4o as the model behind both, provided through Microsoft's Azure. ORG's January paper and WIRED give GPT-4 instead, and the Home Office's own public documents do not name the model.

ACS was fully rolled out in April 2026. Based on material ORG obtained through parliamentary questions and freedom of information requests, Home Office subject matter experts have since checked 203 summaries. In the first round, 45 were judged unusable (22%), and another 19 were flagged and sent for a second-round review. Of 56 summaries reviewed in the second round, 30 were unusable, which the report counts as 15% using 203 as the denominator (30÷203 is about 14.8%, by our calculation), and the report does not explain why the 19 and the 56 do not match.

ORG compares these two figures with the 9% in the Home Office's 2025 pilot evaluation. That figure was the share of summaries removed for inaccuracy or missing information after technical specialists reviewed every summary during the pilot, a different method from the sampling after roll-out. The Register reported that the Home Office did not respond to its questions, and we found no public Home Office response to these figures.

Using monthly averages from the data protection impact assessment (the assessment required before processing high-risk personal data), ORG estimates that about 70,000 cases were processed between April and July, of which only 0.3% were checked by experts. The people affected are those applying for asylum in the UK, and in August the minister responsible for asylum wrote to an MP that applicants are not specifically told AI has been used in the consideration of their claims. The Home Office's privacy notice states that parts of the processing may involve automation and that complex or adverse decisions are always taken by trained staff, with no mention of the two tools. In a legal opinion commissioned by ORG and signed in February (an opinion obtained by the commissioning party and not tested in court), three lawyers concluded that not telling applicants is "likely to be unlawful", and we found no public Home Office response to the opinion.

Perspective

In the Home Office response quoted by WIRED, asylum decisions are made by trained caseworkers, and the tools cannot be used in isolation and do not make decisions on claims. The previous minister's June letter adds that the summary is only one of the documents decision-makers read before deciding, and that they must still read all documents in the case. The Home Office has also committed to publishing updated evaluations.

The 2025 pilot evaluation found that the group using ACS reviewed a transcript 23 minutes faster than the comparison group (a saving of about 32%). By the evaluation's own account, these were small-scale pilots and further evaluation would be needed before a full roll-out. ORG argues that if decision-makers still have to read every document, those time savings cannot be achieved.

Using figures obtained through freedom of information requests, ORG reports that since April 2026 the monthly number of initial substantive decisions has fallen by about 42% compared with the same period a year earlier, and the mean time from interview to decision has risen by 29%. The Register's report notes that these figures do not prove AI caused the change, and that they show little sign of the productivity gains ministers promised. ORG calls for an outright ban on AI in asylum applications, and its report recommends first suspending both tools and commissioning an independent audit.

The two sides describe what the tools do differently. The minister's August letter states that the tools summarise information provided by the claimant and do not generate new content. ORG's report, by contrast, describes both tools as generating new text that can filter, reframe or omit facts relevant to a decision.

ORG's report describes a large language model's summary as newly generated sentences, and the minister's August letter does not explain what "new content" covers. In the pilot evaluation, ACS summaries deliberately did not mark where each point came from in the transcript, to reduce the pilot's technical complexity, and this could be revisited later. Without reference numbers, decision-makers cannot easily check a summary against the original wording of the transcript (our inference), and adding reference numbers to the summaries is among the plans in the minister's letter.

The previous minister's June letter states that experts dip-sample ACS output every week, and the Home Office also told WIRED that it regularly monitors the tools' performance. In ORG's report, the sampling form scores whether a summary omits the most relevant facts, but that criterion is not one of the conditions for failing a summary. On that basis, ORG considers that a summary missing important facts could still pass as long as the rest is accurate, and we have not seen a Home Office explanation of the scoring criteria.

Whether applicants can see the summaries is a separate question. ORG's report states that applicants cannot see or correct the AI summaries of their own interviews. The minister's August letter mentions that applicants can amend the interview record within an agreed period after the interview, and that ACS is used after that period to help decision-makers re-familiarise themselves with the case, but the letter does not say whether applicants can see or correct the summaries.

The UK government runs a set of transparency records through which public bodies publish their algorithmic tools (Algorithmic Transparency Records). When we checked on 29 October there were 154 records, and neither of the Home Office's two is for these tools. The minister's letter explains that both tools were developed before these records became mandatory, and their records are now being completed. ORG's request for the prompts used by the tools was refused, partly on cyber security grounds, and the equality impact assessment (a document assessing a policy's effects on different groups), which the minister said would be published in August, was still unpublished as of 12 October.

In June, the Independent Chief Inspector of Borders and Immigration, who oversees the Home Office's border and immigration functions, published an inspection of asylum casework from June to December 2025, finding that the effectiveness and reliability of APS were varied. In the inspection report, inspectors also found evidence of AI tools being used inappropriately in decision letters and interviews, including Microsoft's AI assistant Copilot, which the Home Office does not permit in asylum interviews and decision-making, and ChatGPT. Managers of the asylum decision-making units told inspectors that these instances had been addressed. WIRED reports that the inspectorate asked ORG for more information to help set next year's inspection topics, but stated that it cannot confirm which topics are under consideration before they are published.

With interview records processed in Azure, who might read an applicant's account depends on the terms of the service (our inference). Microsoft's documentation states that customers' prompts (the text sent to the model) and outputs are not available to OpenAI and are not used to improve the models. The service still runs abuse monitoring to detect misuse, under which prompts and outputs flagged by the system may be reviewed by authorised Microsoft employees, while customers approved for modified abuse monitoring are exempt from this storage and human review.

Microsoft's documentation also lists several deployment types with different data processing locations, and prompts and responses for the Global type may be processed in any geography where the model is deployed. As of 29 October, we could find no explanation from the Home Office itself of which deployment type it uses or whether it has modified abuse monitoring.

Taiwan's Executive Yuan (Taiwan's cabinet) issued its reference guidelines on the use of generative AI by the Executive Yuan and its subordinate agencies (行政院及所屬機關（構）使用生成式AI參考指引) in 2023. They apply to the Executive Yuan and its agencies and are framed as reference guidelines. Under item 5, agencies may not use unverified output to take administrative action directly or as the sole basis for an official decision, which is close to the Home Office's position that the tools cannot be used in isolation. Under item 6, agencies should make appropriate disclosure when using generative AI as an assistive tool, with no wording on to whom, whereas the UK minister's letter states that applicants are not specifically told.

Under Article 19 of Taiwan's AI Basic Act (人工智慧基本法), promulgated in January 2026, the government must also carry out a risk assessment when using AI in its work. We, a community based in Taiwan, found no rule in Taiwan specific to the use of AI in immigration or asylum procedures.

South Korea draws the line in statute rather than in guidance, and under Article 20 of the General Act on Public Administration, an administrative authority may impose a disposition using a fully automated system, including systems that employ AI, except for dispositions imposed at its discretion. We read the article in the English translation by the Korea Legislation Research Institute and did not check the Korean original, which secondary sources describe as adding wording to the effect of "as prescribed by statute". We have not read the primary text of Article 37-2 of Korea's Personal Information Protection Act, the provision on automated decisions, so whether Korean law requires telling the person affected that AI was involved remains unverified. We found no provision specific to asylum procedures.

In mainland China, the Guidelines for the Deployment and Application of Large AI Models in Government Affairs were issued by the Office of the Central Cyberspace Affairs Commission and the National Development and Reform Commission, and published on the Cyberspace Administration of China's website on 10 October 2025. The opening of the guidelines describes them as providing orientation and a basic reference for government departments at all levels. Under the guidelines, large models are to stay in an "assistive" role, and applications must display a risk warning in a prominent place on the interface, state the service's limitations and label model output, which is aimed at people using the interface rather than at those affected by a decision. The guidelines contain no wording against using model output as the sole basis for a decision.

China's Personal Information Protection Law applies to state organs under Article 33, and under Article 24, individuals may ask for an explanation of a decision made through automated decision-making that significantly affects their rights and interests, and may refuse a decision made solely by automated means. The provision covers automated decision-making in general and does not mention generative AI, and we did not look into how state organs apply it in practice. Under Article 36, personal information handled by state organs must be stored within China, with a security assessment required if it genuinely needs to be provided abroad. We found no rule in mainland China specific to the use of generative AI in asylum or immigration reviews.

Japan's guideline on the procurement and use of generative AI in government (DS-920), decided in May 2025 by the executive board of the Digital Society Promotion Council, is labelled "Normative", a set of rules to be complied with in developing and managing government information systems. The guideline contains no explicit ban on relying solely on AI output and handles the issue through risk assessment instead. Each ministry's Chief AI Officer uses a high-risk assessment sheet on which work with a major effect on people's fundamental rights or safety, and output used without a government employee's judgement, both raise the risk level, and cases judged high-risk must be reported to an advisory board on advanced AI use.

We found no provision in the guideline on telling people affected by an administrative decision that AI was involved. The provisions above are from the first edition of May 2025. A version 2.0 was adopted in June 2026 according to press reports, and we have not read it.

People dealing with agencies under Taiwan's Executive Yuan can ask the agency handling their case whether generative AI is used on it (our suggestion). Item 6 contains no wording on to whom disclosure is owed or how much must be disclosed, so asking does not guarantee a complete answer. Beyond that, readers do not need to change their habits because of this.

When comparing how governments use AI on individual cases, readers can look at whether the people concerned are told, whether they can see and correct what the AI produced, whether the sampling rate and criteria are published, and whether omitted facts count as errors. The intended effect of the policy matters too, including whether the time saved outweighs the cost of errors, who carries out the evaluation, and whether the results are published.