The UK Uses AI to Summarize Asylum Interviews. A Watchdog May Now Look at How Well It Works

The UK’s borders watchdog is weighing whether to inspect how the Home Office uses generative AI to help decide asylum claims. The Independent Chief Inspector of Borders and Immigration (ICIBI) has asked the digital rights group Open Rights Group (ORG) for more information about two AI tools as it plans next year’s inspection program, according to reporting by WIRED.

The ICIBI would not confirm which topics are under consideration before it publishes its priorities. It said it is aware of concerns about AI and pointed to its own finding last year that one of the tools’ “effectiveness and reliability were varied.” Whatever the inspector decides, the case already offers a clear lesson for any organization using AI to summarize records about people: if you cannot explain the tool, list it and let people challenge it, you will be asked why not.

The two tools

According to documents obtained by ORG, the Home Office uses two tools built on OpenAI’s 2023 GPT-4 model and Microsoft’s AI platform:

• Asylum Case Summarisation (ACS) summarizes transcripts of interviews in which applicants describe persecution and why they fear returning home.
• Asylum Policy Search (APS) works as a search assistant that finds and summarizes country policy information, which caseworkers use to judge whether an applicant faces a real risk of harm.

The Home Office began piloting them in late 2024 while trying to clear a backlog of around 100,000 asylum cases. Government estimates suggested that used together, they could save about an hour per case. The department has since halved the number of people waiting for an initial decision in the year to June, although appeals against decisions have risen sharply.

What critics found

ORG’s findings come from a six-month investigation that involved 11 parliamentary questions, two ministerial letters and 10 public records requests. The group says:

• the share of ACS summaries judged not fit for use, often because of factual errors or misrepresentations, is significantly higher than the 9% error rate published from the pilot evaluation;
• there is no quality control for summaries produced by APS;
• applicants are not told when the tools are used in their cases; and
• neither tool appears on the government’s public algorithmic transparency repository, even though the government’s AI Playbook expects departments to be transparent about algorithmic tools used in decision-making.

The Public Law Project, which has investigated separately, says these are two of 11 high-risk AI systems it has found in use across government but missing from the repository. In July, a UK judge accused the Home Office of citing an “AI hallucinated” document when refusing an asylum claim from a Moroccan woman and her child. Researchers and Labour MP Bell Ribeiro-Addy have also raised concerns about cultural bias in GPT models and the loss of nuance in translation-heavy interviews.

These are findings and concerns from advocacy groups and lawmakers, not conclusions of a regulator or court. The Home Office disputes the characterization.

The Home Office’s position

“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,” a Home Office spokesperson said. The department says it monitors the tools’ performance alongside its quality-assurance processes and has committed to publishing updated evaluations. Its privacy notice states that parts of the decision-making process may involve automation. OpenAI said it continues to work on accuracy and noted that GPT-4 will be retired in January.

Why “a human decides” is not the end of the analysis

The Home Office’s defense rests on a familiar argument: the AI assists, a human decides. Under UK data protection law that matters, because the strictest rules on automated decision-making apply to decisions based solely on automated processing. But it does not settle every question.

1. Summaries shape decisions. If a caseworker reads an AI summary instead of the full transcript, errors in the summary can flow into the decision even with a human in the loop.
2. Accuracy is a legal principle. UK GDPR requires personal data to be accurate. A summary that misstates what an applicant said is inaccurate personal data about that person.
3. People must be able to exercise their rights. A legal opinion from Cloisters barristers, commissioned by ORG, argues that applicants’ inability to contest an incorrect AI summary may breach UK data rights law.
4. Risk assessment is expected. High-risk processing of special category data, with vulnerable people, calls for a data protection impact assessment that tests the tool, not just describes it.

Both ORG and the Public Law Project have sent their findings to the UK’s data regulator, the Information Commission, which is investigating the Home Office’s refusal to tell ORG how staff use and were trained to use the two tools. The regulator has said it prefers working with public bodies to change practice over large fines, as its public-sector enforcement approach sets out, so any outcome is more likely to be orders and recommendations than a penalty. Individuals who believe their data was mishandled can also complain directly, using the process described in our guide to the data protection complaints framework.

Lessons for organizations outside government

Private companies use the same pattern every day: AI that summarizes support calls, claims files, HR investigations or medical notes for a human who then decides. The asylum case shows where scrutiny lands.

• Measure the error rate in production, not just the pilot. ORG’s central claim is that real-world failure rates exceed the published pilot figure.
• Tell people. If AI processes someone’s personal data in a decision that affects them, say so in the notice and at the point of use.
• Keep the source. Make sure reviewers can and do check the original record, and log when they rely on the summary alone.
• Plan for model retirement. GPT-4’s retirement in January means the tool will change. Re-test whenever the underlying model does.

Our earlier analysis argued that the DPIA is broken when it becomes a paperwork exercise. Cases like this show why: an assessment is not proof that a tool works. For organizations that both build and deploy AI, our guide to DPIA considerations for dual-role AI providers covers the overlap. Regulators in the United States are applying similar logic; see Connecticut’s algorithmic accountability blueprint.

How Captain Compliance can help

Captain Compliance’s UK GDPR tools support the records this kind of scrutiny demands: data protection impact assessments, records of processing, privacy notices that can be updated as tools change, with every version kept, and a request portal for handling access and rectification requests. Those records help an organization show what it assessed and what it told people. They do not prove that an AI summary is accurate, and we do not guarantee compliance. Testing the tool remains the controller’s job.

Frequently asked questions

Has the ICIBI opened an investigation?

Not publicly. It has asked ORG for information while planning next year’s inspections and would not confirm which topics it is considering.

Do the AI tools decide asylum claims?

The Home Office says no. It says trained caseworkers make decisions and the tools only help them navigate interview records and policy information.

What is the published error rate?

The pilot evaluation reported a 9% error rate for ACS summaries. ORG says the real share of unusable summaries is significantly higher.

Is the Information Commission involved?

Yes. It is investigating the Home Office’s refusal to tell ORG how staff use and were trained on the two tools, and it has received both groups’ findings.

Do UK automated decision rules apply when a human decides?

The strictest rules apply to decisions based solely on automated processing. Other duties, including accuracy, transparency, fairness and data subject rights, apply whenever AI processes personal data.

The repository with a gap in it

The UK built a public register so people could see which algorithms help decide their cases. The two tools at the center of this dispute are not on it. Whether or not the inspector takes up the case, the empty space where those entries should be has become the strongest argument critics have.