英国边境监察机构或调查内政部使用 OpenAI 工具处理庇护申请

A UK government watchdog is considering an investigation into the Home Office’s use of generative AI tools to process asylum applications, WIRED has learned. The potential probe follows lawmakers and researchers publicly questioning the tools’ accuracy and the government’s secrecy around their use.

The Independent Chief Inspector of Borders and Immigration (ICIBI), which monitors the Home Office’s management of borders, immigration, and customs, has asked digital rights nonprofit Open Rights Group (ORG) to share more information on the tools as it draws up plans for the topics it will inspect next year, campaigner Sara Alsherif tells WIRED. ORG’s findings come from a six-month investigation into the tools, which involved 11 parliamentary questions, two ministerial letters, and 10 public records requests that attempt to pry details from the Home Office, which has withheld details on the tools.

The ICIBI tells WIRED it cannot confirm what topics are under active consideration before it publishes its annual inspection priorities, but it says it is aware of concerns about the use of AI tools and pointed to its findings last year that one of the tool’s “effectiveness and reliability were varied.”

“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 tells WIRED in an email. “We regularly monitor the performance of these tools alongside our established quality-assurance processes, and have already committed to publishing updated evaluations on their roll out.”

The Home Office uses the Asylum Case Summarisation (ACS) and Asylum Policy Search (APS) tools, built on OpenAI’s 2023 GPT-4 large language model and Microsoft Foundry, to summarize transcripts of interviews with asylum applicants and information about their countries of origin, according to documents obtained by ORG. The immigration and law enforcement authority started piloting these tools in late 2024 as the Labour government tried to clear a backlog of around 100,000 asylum cases, while staying on top of the tens of thousands of new applications the UK receives each year. While the department has made significant progress, halving the number of people waiting on an initial decision in the year through June, the number of appeals against decisions has ballooned.

Researchers are concerned about the accuracy of the AI tools–especially to decide something as high-stakes as whether to send someone back to a country where they fear persecution. ORG found that the percentage of summaries generated by the ACS deemed not fit for use—often because they contain factual errors or misrepresentations—is significantly higher than the published 9 percent error rate calculated in the pilot evaluation. There is no quality control for the summaries produced by the APS, it also found.

Asylum decision-makers use the tools to save time when reviewing asylum cases; when used together, the pilot showed the tools could save around an hour per case, according to government estimates. Caseworkers use the ACS to speed up their review of transcripts of interviews where asylum seekers explain how they were persecuted in their country and why they are afraid to return. They use the APS as a search assistant to find and summarize country policy information, which informs them whether the applicant faces a genuine risk of harm in their home country. In July, a UK judge accused the Home Office of citing an “AI hallucinated” document to refuse a Moroccan woman and her child’s asylum claim on the basis that her home country would be safe for her to return to.

“We’re dealing with people’s lives. We’re talking about people that have faced some of the worst circumstances, and we’re meant to be making decisions based on facts and on understanding harm,” Labour backbench MP Bell Ribeiro-Addy tells WIRED. “I just don’t think it’s appropriate that machines are doing that.”

A Home Office spokesperson says, “Use of Artificial Intelligence 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.”

Ribeiro-Addy, Alsherif, and the Public Law Project, which has separately been investigating the tools, also worry that well-documented “cultural bias” in GPT models could lead to language barriers and the loss of cultural nuances in the summaries, which could in turn result in discriminatory decisions. An OpenAI spokesperson highlighted to WIRED the company’s ongoing work to improve its models’ accuracy and pointed out that the company will be retiring GPT-4 in January.

In the absence of regulation, the UK’s “AI Playbook” mandates that departments must be transparent about the algorithmic tools they use in decision-making processes and make information on them available to the public, including listing those tools on a public repository. The Home Office’s case summarization and policy search tools are not on that list.

Asylum seekers are also not told when these AI tools are used in their case. “Migrant organizations, legal practitioners, not many of them know about the actual use of this tool,” says Alsherif, who leads ORG’s migrant digital justice program. “So we need to let them know about this, let them know how it works, so they can challenge it.” Not being able to contest an incorrect AI summary of their personal information may breach UK data rights laws, Cloisters lawyers wrote in an independent legal opinion commissioned by ORG. The Home Office states that parts of the decision-making process may involve automation in its privacy notice.

Both groups have sent their findings to UK data regulator, the Information Commission, which is currently investigating the Home Office’s refusal to tell ORG how staff use and were trained to use two AI tools.

Public Law Project senior research fellow Caroline Selman also notes that the ACS and APS tools are just two of the 11 high-risk AI systems that her team has found being used across government but absent from the public repository. She’s concerned there may be more deployed across government that even researchers, let alone the individuals they are being used on, are unaware of.

“Other tools are being used but not being publicized at all,” says Selman, “and you can’t even begin to ask questions about things if you don’t know that they’re being used.”