英国内政部最新尝试使用人工智能来加速难民申请裁决的工作遇到了一个尴尬的挫折：一份新报告显示，经专家检查的面谈总结中，有五分之一以上未能通过准确率测试。
The Home Office's latest attempt to speed up asylum decisions with AI has hit an awkward snag: a new report says more than one in five interview summaries checked by experts failed accuracy tests.
开放权利组织（ORG）发布的一份长达20页的调查报告称，尽管对人工智能工具的准确率、监督以及承诺的效率提升存在担忧，内政部仍在执意推进这些工具。这些发现基于信息自由请求以及与部长的信函往来，涉及内政部案卷处理人员使用的两个系统：用于总结难民面谈的“难民案卷总结”（ACS）系统，以及帮助官员检索国家政策信息的聊天式工具“难民政策搜索”（Asylum Policy Search）。两者均通过微软的Azure AI Foundry使用OpenAI的GPT-4o模型。
A 20-page investigation by the Open Rights Group (ORG) claims the department is pressing ahead with AI tools despite concerns about their accuracy, oversight, and promised efficiency gains. The findings, based on Freedom of Information requests and correspondence with ministers, concern two systems used by Home Office caseworkers: Asylum Case Summarisation (ACS), which summarizes asylum interviews, and Asylum Policy Search, a chat-style tool helping officials search country policy information. Both use OpenAI's GPT-4o model through Microsoft's Azure AI Foundry.
政府坚称人类仍然对难民决定负有责任，但开放权利组织质疑案卷处理人员能够多么可靠地发现人工智能生成信息中的错误。该活动集团获得的数据显示，自该系统于2026年4月推出以来，在人工智能生成的203份面谈总结中，有45份未通过最初的专家检查。这相当于22%的不合格率，而最初试点期间记录的拒绝率为9%。另有19份总结被标记需要进一步审查。
The government insists humans remain responsible for asylum decisions, but ORG questions how reliably caseworkers can spot mistakes in AI-generated information. Figures obtained by the campaign group show that 45 of 203 AI-generated interview summaries failed initial expert checks after the system was rolled out in April 2026. That's a 22 percent failure rate, compared with the 9 percent rejection rate recorded during the original pilot. Another 19 summaries were flagged for further review.
这些数据还记录了56次第二阶段审查，其中30次未能达到标准。这些错误可能包括事实不准确以及对难民在面谈中所言内容的歪曲。开放权利组织还发现，只要所包含的信息准确，即使总结遗漏了重要细节，也可以通过检查。
The figures also record 56 second-stage reviews, of which 30 failed to meet the required standard. The errors can include factual inaccuracies and misrepresentations of what asylum seekers said during interviews. ORG also found that summaries could pass the checks even if they omitted important details, as long as the information they did include was accurate.
尽管有这些结果，但接受检查的总结相对较少。根据内政部预测的每月17,515起案件量，开放权利组织估计，内政部专家在4月至7月期间仅检查了0.3%的总结。开放权利组织还质疑这些工具是否节省了多少时间。该组织表示，自2026年4月以来，每月初步实质性难民裁决的平均数量与2025年同期相比下降了约42%，而从面谈到做出裁决的平均时间则增加了29%。虽然这些数据并未证明人工智能导致了效率放缓，但它们几乎没有提供部长们所承诺的生产力提升的证据。
Despite those results, relatively few summaries have been examined. ORG estimates that Home Office experts checked only 0.3 percent between April and July, based on the department's projected monthly caseload of 17,515. ORG also questioned whether the tools were saving much time. It says the average monthly number of initial substantive asylum decisions since April 2026 fell about 42 percent compared with the same period in 2025, while the mean time from interview to decision rose 29 percent. While those figures don't prove that AI caused the slowdown, they offer little evidence of the productivity gains ministers have promised.
对另一个“难民政策搜索”工具的监督也显得缺乏。尽管此前曾告诉议员检查仍在继续，但内政部在8月确认专家测试已经停止。开放权利组织表示，并未产出测试结果的总结。
Oversight of the separate Asylum Policy Search tool also appears lacking. The Home Office confirmed in August that expert testing had stopped, despite previously telling MPs that checks were continuing. ORG says no summary of the testing results was produced.
与此同时，难民申请人并没有被明确告知其案件中使用了人工智能，也无法查阅或更正从其面谈中生成的总结。开放权利组织还对内政部关于人类监督的主张提出了质疑。该部门自己的数据保护评估显示，难民案卷总结系统将以前由技术专家完成的工作自动化了，这些技术专家在将难民案件转交给决策者之前会对其进行审查和总结。开放权利组织警告称，该工具可能会取消经验丰富的审查层级，而不仅仅是节省员工时间。
Meanwhile, asylum applicants are not specifically told when AI has been used in their cases and cannot inspect or correct the summaries generated from their interviews. ORG also questions the Home Office's claims of human oversight. The department's own data protection assessment says ACS automates work previously done by technical specialists, who reviewed and summarized asylum cases before passing them to decision-makers. ORG warns the tool could remove an experienced layer of scrutiny, rather than simply save staff time.
该组织呼吁暂停这两个工具并进行独立审计，并认为目前的保障措施不足以用于决定某人能否留在英国的决策。
The group is calling for both tools to be suspended and independently audited, arguing that the current safeguards are inadequate for decisions that determine whether someone can remain in the UK.
这并非内政部的技术计划首次受到抨击。今年6月，60多个权利组织敦促部长们放弃对难民儿童进行人工智能年龄评估的计划，并警告称该技术在关键的16至18岁年龄段不可靠。
It's not the first time the Home Office's technology plans have come under fire. In June, more than 60 rights groups urged ministers to abandon plans for AI-powered age estimation of asylum-seeking children, warning that the technology was unreliable around the crucial 16-to-18-year-old age range.
同月晚些时候，一项检查发现，该部门长期延误的Atlas移民系统未能向案卷处理人员提供难民上诉的反馈，迫使员工依赖变通方法。内政部未回应相关问题。
Later that month, an inspection found that the department's long-delayed Atlas immigration system was failing to provide caseworkers with feedback from asylum appeals, forcing staff to rely on workarounds. The Home Office did not respond to The's questions. ®