UK Home Office May Face Investigation Over Use of OpenAI Tech

WIRED reports that the UK's Independent Chief Inspector of Borders and Immigration is gathering information that could lead to an inspection of the Home Office's generative AI tools for asylum cases. The inspector would not confirm its plans but said it knows about the concerns and noted its own earlier finding that one tool's effectiveness and reliability were varied.

The tools, Asylum Case Summarisation and Asylum Policy Search, run on OpenAI's 2023 GPT-4 model and Microsoft Foundry. Caseworkers use them to summarize interview transcripts in which applicants explain why they fear returning home, and to search country policy information. The government estimated the pair could save about an hour per case. These are decisions about whether someone may be sent back to a place where they fear persecution.

We support using AI to cut real backlogs. The UK faced roughly 100,000 pending asylum cases when the pilots began in late 2024, and halving the number waiting on an initial decision in the year through June is meaningful for people stuck in limbo. Assistive tools that help trained caseworkers find information faster can be a public good.

But the evidence here points the other way on trust. Open Rights Group, after a six-month investigation involving parliamentary questions, ministerial letters, and records requests, found the share of case summaries judged not fit for use is significantly higher than the 9 percent error rate the pilot published. It found no quality control on policy search summaries at all. In July, a UK judge accused the Home Office of relying on an AI-hallucinated document to refuse a Moroccan woman and her child. Appeals, meanwhile, have ballooned.

The Home Office says caseworkers make decisions, the tools only assist, and performance is monitored. I take that seriously, and I also know how assistance works in practice under backlog pressure. When a summary is the fastest path through a long interview, the summary becomes the record. Errors and lost cultural nuance in the summary, which campaigners and MP Bell Ribeiro-Addy warn about, then quietly shape the outcome.

The secrecy is what turns a quality problem into an accountability problem. The UK's own AI Playbook asks departments to list algorithmic decision tools in a public repository. These tools are not on it. Applicants are not told when the tools touch their case, so they cannot challenge a wrong summary, which lawyers at Cloisters say may breach data rights. The Public Law Project says these are two of 11 high-risk systems it found missing from that list.

If this scales, we get a state that processes vulnerable people with tools neither they nor their lawyers know exist. Our read: an inspection is warranted. The fix is not banning assistance. It is disclosure, independent error audits, and a right to see and contest the summary.

The UK data regulator, the Information Commission, is already investigating the Home Office's refusal to tell Open Rights Group how staff use and were trained on two AI tools. OpenAI told WIRED it is working on model accuracy and will retire GPT-4 in January, which raises its own question about what the department runs next.