UK Home Office May Face Investigation Over Use of OpenAI Tools
出典: Wired
A UK government watchdog is considering launching an investigation into the Home Office's use of generative AI tools to process asylum applications. The potential probe follows growing concern from lawmakers and researchers about the accuracy of the tools and the government's secrecy surrounding their deployment. The Independent Chief Inspector of Borders and Immigration (ICIBI) has reached out to the digital rights group Open Rights Group (ORG) for more information as it draws up its inspection priorities for the coming year.
ORG's findings stem from a six-month investigation that included 11 parliamentary questions, two ministerial letters, and 10 public records requests. Despite these efforts, the Home Office has withheld significant details about the tools. The ICIBI acknowledged awareness of concerns around AI use in immigration decisions and referenced its own prior finding that one tool's 'effectiveness and reliability were varied.' A Home Office spokesperson insisted that AI only assists caseworkers and does not make decisions on asylum claims.
The Home Office currently uses two tools—the Asylum Case Summarisation (ACS) and the Asylum Policy Search (APS)—built on OpenAI's GPT-4 model and Microsoft Foundry. These tools summarize interview transcripts and country-of-origin information for asylum applicants. The department began piloting them in late 2024 as part of a broader effort by the Labour government to clear a backlog of roughly 100,000 asylum cases. While the backlog has been halved, the number of appeals against decisions has risen sharply.
Researchers have raised serious concerns about the accuracy of the AI tools, particularly given the high-stakes nature of asylum decisions, where errors could mean returning someone to a country where they face persecution. ORG found that the actual error rate in ACS-generated summaries—often containing factual errors or misrepresentations—is significantly higher than the 9 percent figure published in the pilot evaluation. Additionally, no quality control mechanism exists for summaries produced by the APS tool.
The investigation highlights a broader tension between the government's drive to use AI to increase efficiency and the need for transparency and accountability in life-altering decisions. Digital rights advocates argue that the Home Office's reluctance to disclose details about these tools undermines public trust and due process for asylum seekers. If the ICIBI proceeds with a formal inquiry, it could set an important precedent for how AI systems are scrutinized within sensitive areas of government.
日本語訳
英国政府の監視機関が、難民申請の処理に生成AI ツールを使用している内務省の調査を検討している。この動きは、ツールの精度や政府の情報秘匿に対する議員や研究者からの批判を受けたものだ。移民・国境管理を監視する独立機関ICIBIが、デジタル権利団体「オープン・ライツ・グループ（ORG）」に情報提供を求め、来年の検査優先事項の策定に活用しようとしている。
ORGの調査結果は、6か月にわたる取り組みから得られたもので、11件の議会質問、2通の大臣書簡、10件の情報公開請求が含まれる。しかし内務省は重要な情報の開示を拒んでいる。ICIBIはAIの使用に関する懸念を把握していると認め、過去の調査でもツールの「有効性と信頼性にばらつきがある」と指摘していた。内務省の広報担当者は、AIはあくまで補助ツールであり、審査官に代わって判断を下すものではないと強調した。
内務省は現在、OpenAIのGPT-4とマイクロソフト・ファウンドリーを基盤とした2つのツール——「難民事例要約（ACS）」と「難民政策検索（APS）」——を使用している。これらは申請者の面接記録や出身国情報を要約するためのものだ。2024年後半に試験的な運用が始まったのは、労働党政権が約10万件の難民申請の滞留解消を目指したためで、積み残しは半減したが、決定への異議申し立ては急増している。
研究者たちは特に、AIツールの精度に深刻な懸念を示している。難民申請の審査は、迫害を恐れる人物を出身国に送り返す可能性がある、極めて重大な決定だからだ。ORGの調査では、ACSが生成する要約の実際のエラー率——事実誤認や誤った表現を含むケースが多い——が、パイロット評価で公表された9%を大幅に上回ることが判明した。さらにAPSには品質管理の仕組みが存在しないことも明らかになった。
この問題は、効率向上を目指す政府のAI活用推進と、人生を左右する決定における透明性・説明責任の確保との間の根本的な緊張関係を浮き彫りにしている。デジタル権利の擁護者たちは、内務省がツールの詳細を開示しないことが、公的信頼と難民申請者の適正手続きを損なっていると訴えている。ICIBIが正式な調査に踏み切れば、政府の機密性の高い分野でのAIシステムの精査に重要な先例を打ち立てることになる。
重要単語・フレーズ
- watchdognoun
An independent organization that monitors and oversees the activities of companies or government bodies to ensure they act fairly and legally.
和訳: 監視機関、監視団体
The financial watchdog launched an inquiry into the bank's lending practices after numerous customer complaints.
- pry details fromphrase
To obtain information that someone is reluctant to share, often through persistent effort or pressure.
和訳: （情報を）無理やり引き出す
Journalists spent weeks trying to pry details from the corporation about its data breach.
- withheldverb
Past tense of 'withhold'; to deliberately refuse to give or share information or something that is wanted.
和訳: （情報などを）隠す、開示しない
The agency was criticized for having withheld key evidence from the public inquiry.
- backlognoun
An accumulation of unfinished work or unprocessed cases that have built up over time.
和訳: （処理待ちの）案件の積み残し、滞留
The court system is struggling to clear a massive backlog of cases that built up during the pandemic.
- high-stakesadjective
Involving serious consequences or significant risks, where the outcome matters greatly.
和訳: 重大な結果をもたらす、リスクの高い
Diagnosing a rare disease is a high-stakes process that demands both accuracy and speed.
- misrepresentationnoun
A false or inaccurate account of something, whether intentional or not, that gives a misleading impression.
和訳: 虚偽表示、誤った表現
The defendant claimed that the advertising contained deliberate misrepresentations about the product's safety record.
- due processnoun
The legal requirement that all individuals receive fair treatment and their legal rights are respected in judicial or administrative proceedings.
和訳: 適正手続き
Critics argued that deporting individuals without a hearing violated their right to due process.
- precedentnoun
An earlier decision or action that serves as an example or rule to follow in similar future situations.
和訳: 先例、前例
The court's ruling set a precedent that would affect how online privacy cases are handled for years to come.
- accountabilitynoun
The obligation of an individual or organization to accept responsibility for their actions and be open to scrutiny.
和訳: 説明責任
Increased accountability in public institutions helps maintain citizens' trust in government.