英国 Ada Lovelace 協会 英国のAI規制、AIリスクなど4つの報告書
こんにちは、丸山満彦です。
英国 Ada Lovelace 協会は最近AIに関する報告書を公表していますが、今回4本の報告書を公表していますね。。。
● Ada Lovelace Institute - Library - Reports
・2023.07.18
Regulating AI in the UK | 英国におけるAIの規制 |
Policy briefing: Regulating AI in the UK | 政策ブリーフィング:英国におけるAI規制 |
AI risk: Ensuring effective assessment and mitigation across the AI lifecycle | AIリスク:AIのライフサイクルを通じた効果的な評価と低減の確保 |
Keeping an eye on AI | AIから目を離さない |
・2023.07.18 Regulating AI in the UK
・[PDF]
Regulating AI in the UK | 英国におけるAIの規制 |
About this report | 本報告書について |
This report is aimed at policymakers, regulators, journalists, AI practitioners in industry and academia, civil society organisations, and anyone else who is interested in understanding how AI can be regulated in the UK for the benefit of people and society. | 本報告書は、政策立案者、規制当局者、ジャーナリスト、産業界や学界のAI実務者、市民社会組織、その他、人々と社会の利益のために英国でAIをどのように規制できるかを理解することに関心のあるすべての人を対象としている。 |
It contextualises and summarises the UK’s current plans for AI regulation, and sets out recommendations for the Government and the Foundation Model Taskforce. | 本書は、AI規制に関する英国の現在の計画を整理・要約し、ガバナンスとファウンデーション・モデル・タスクフォースに対する提言を示している。 |
Our recommendations are based on extensive desk research, two expert roundtables, independent legal analysis and the results of a nationally representative survey. They also build on previous Ada Lovelace Institute research, including the 2021 report Regulate to innovate1 and extensive analysis and commentary on the EU AI Act.2 For more information on our evidence base, see the sections below on ‘Methodology’ and ‘Further reading’. | 我々の提言は、広範な机上調査、2回の専門家懇談会、独立した法的分析、全国代表者調査の結果に基づいている。また、2021年の報告書「Regulate to innovate」1 やEUのAI法に関する広範な分析と解説2 を含む、エイダ・ラブレス研究所の過去の調査にも基づいている。証拠ベースの詳細については、以下の「方法論」と「参考文献」のセクションを参照のこと。 |
If you would like more information on this report, or if you would like to discuss implementing our recommendations, please contact our policy research team at [mail]. | 本報告書に関する詳細、または提言の実施に関するご相談は、政策研究チーム [mail] までご連絡いただきたい。 |
Executive summary | 要旨 |
It seems as if discussions about artificial intelligence (AI) are everywhere right now – new and emerging uses of AI technologies are appearing across different sectors and are also implicit in every conversation about present and future societies. | 今、人工知能(AI)に関する議論は至る所で行われているように思われる。AI技術の新たな用途や新興の用途が様々な分野で登場し、現在と未来の社会に関するあらゆる会話にも暗黙の了解となっている。 |
The UK Government has laid out its ambition to make the UK an ‘AI superpower’, leveraging the development and proliferation of AI technologies to benefit the UK’s society and economy, and hosting a global summit in autumn 2023. | 英国政府は、英国を「AI大国」にするという野望を打ち出し、AI技術の開発と普及を活用して英国の社会と経済に利益をもたらし、2023年秋にはグローバル・サミットを開催するとしている。 |
This ambition will only materialise with effective domestic regulation, which will provide the platform for the UK’s future AI economy | この野望は、効果的な国内規制があって初めて実現するものであり、それが英国の将来のAI経済のプラットフォームを提供することになる |
‘Regulating AI’ means addressing issues that could harm public trust in AI and the institutions using them, such as data-driven or algorithmic social scoring, biometric identification and the use of AI systems in law enforcement, education and employment. | 「AIを規制する」とは、データ駆動型またはアルゴリズムによる社会的スコアリング、生体データによる識別、法執行、教育、雇用におけるAIシステムの使用など、AIとそれを使用する機構に対する国民の信頼を損ないかねない問題に対処することを意味する。 |
Regulation will need to ensure that AI systems are trustworthy, that AI risks are mitigated, and that those developing, deploying and using AI technologies can be held accountable when things go wrong. | 規制は、AIシステムが信頼に足るものであること、AIのリスクが低減されていること、そしてAI技術を開発・導入・使用する者が、問題が発生した際に説明責任を果たせることを保証する必要がある。 |
The UK’s approach to AI regulation | AI規制に対する英国のアプローチ |
While the EU is legislating to implement a rules-based approach to AI governance, the UK is proposing a ‘contextual, sector-based regulatory framework’, anchored in its existing, diffuse network of regulators and laws.3 | EUがAIガバナンスにルールベースのアプローチを導入するための法制化を進めているのに対し、英国は、既存の規制当局と法律の拡散したネットワークに支えられた「文脈に基づく、セクターベースの規制枠組み」を提案している3。 |
The UK approach, set out in the white paper Establishing a pro-innovation approach to AI regulation rests on two main elements: AI principles that existing regulators will be asked to implement, and a set of new ‘central functions’ to support this work. 4 | 白書「Establishing a pro-innovation approach to AI regulation」で示された英国のアプローチは、主に2つの要素から成り立っている: すなわち、既存の規制当局が実施するよう求められるAIの原則と、この作業をサポートする一連の新しい「中央機能」である。 4 |
In addition to these elements, the Data Protection and Digital Information Bill currently under consideration by Parliament is likely to impact significantly on the governance of AI in the UK, as will the £100 million Foundation Model Taskforce and AI Safety Summit convened by the UK Government.5 6 7 | これらの要素に加え、現在議会で審議中のデータ保護とデジタル情報法案は、英国政府によって招集された1億ポンドの財団モデル・タスクフォースとAI安全サミットと同様に、英国におけるAIのガバナンスに大きな影響を与える可能性が高い5 6 7。 |
At the Ada Lovelace Institute we have welcomed the allocation of significant Government resource and attention to AI safety, and its commitment to driving AI safety forward at a global level. It will be important that the definition of ‘safety’ adopted by the Government is an expansive one, reflecting the wide variety of harms arising as AI systems become more capable and embedded in society. | エイダ・ラブレス協会では、政府がAIの安全性に多大な資源と関心を割り当て、世界レベルでAIの安全性を推進するというコミットメントを表明したことを歓迎している。政府が採用する「安全」の定義は、AIシステムがより高性能になり社会に組み込まれるにつれて生じる多種多様な危害を反映した、広範なものであることが重要であろう。 |
It is also unlikely that international agreements will be effective in making AI safer and preventing harm, unless they are underpinned by robust domestic regulatory frameworks that can shape corporate incentives and developer behaviour in particular. The credibility of the UK’s AI leadership aspirations therefore rests on getting the domestic regime right. | また、特に企業のインセンティブや開発者の行動を形成できる強固な国内規制の枠組みに裏打ちされない限り、国際協定がAIをより安全なものにし、危害を防止する上で効果的である可能性は低い。従って、英国がAIのリーダーを目指すことの信頼性は、国内体制を正しく構築できるかどうかにかかっている。 |
Regulating AI in the UK: our recommendations | 英国におけるAIの規制:我々の提言 |
Our recommendations fall into three categories, reflecting our three tests for effective AI regulation in the UK: coverage, capability and urgency. | 我々の提言は、英国における効果的なAI規制のための3つのテスト、すなわち「適用範囲」、「能力」、「緊急性」を反映し、3つのカテゴリーに分類される。 |
Coverage | 適用範囲 |
AI is being deployed and used in every sector but the UK’s diffuse legal and regulatory network for AI currently has significant gaps. Clearer rights and new institutions are needed to ensure that safeguards extend across the economy. | AIはあらゆる分野で導入・利用されつつあるが、AIに関する英国の拡散した法的・規制的ネットワークには現在、大きなギャップがある。セーフガードを経済全体に広げるためには、より明確な権利と新たな機構が必要である。 |
Legal rights and protections | 課題:法的権利と防御 |
New legal analysis shows safeguards for AI-assisted decision-making don’t properly protect people. | 新たな法的分析によると、AIによる意思決定のセーフガードは人々を適切に保護していない。 |
Recommendation 1: Rethink the elements of the Data Protection and Digital Information Bill that are likely to undermine the safe development, deployment and use of AI, such as changes to the accountability framework. | 提言1:データ保護・デジタル情報法案のうち、AIの安全な開発・配備・利用を損なう可能性の高い要素を再考する、 例えば、説明責任の枠組みの変更などである。 |
Recommendation 2: Review the rights and protections provided by existing legislation such as the UK General Data Protection Regulation (GDPR) and the Equality Act 2010 and – where necessary – legislate to introduce new rights and protections for people and groups affected by AI to ensure people can achieve adequate redress. | 提言2:英国一般データ保護規則(GDPR)や2010年平等法(Equality Act 2010)などの既存の法律が提供する権利と保護を見直し、必要な場合には、AIによって影響を受ける人々やグループが適切な救済を受けられるよう、新たな権利と保護を導入するための立法を行う。 |
Recommendation 3: Publish a consolidated statement of the rights and protections that people can expect when interacting with AI-based products and services, and organisations providing them. | 提言3:AIを利用した製品やサービス、そしてそれらを提供する組織と対話する際に人々が期待できる権利と保護について、統合されたステートメントを公表する。 |
Routes to redress | 課題:救済への道 |
Even when legal safeguards are in place, accessing redress can be costly and unrealistic for many affected people. | 法的な保護措置が講じられていても、被害を受けた多くの人々にとって、救済を受けるには費用がかかり、現実的でない場合がある。 |
Recommendation 4: Explore the value of establishing an ‘AI ombudsman’ to support people affected by AI and increase regulators’ visibility of AI harms as they occur. | 提言4:AIの影響を受ける人々を支援し、AIによる被害が発生した際に規制当局の可視性を高めるために、「AIオンブズマン」を設置する価値を検討する。 |
Regulatory gaps | 課題:規制のギャップ |
The Government hasn’t addressed how its proposed AI principles will apply in many sectors. | ガバナンスは、提案されているAIの原則が多くの分野でどのように適用されるかを取り上げていない。 |
Recommendation 5: Set out how the five AI principles will be implemented in domains where there is no specific regulator and/or ‘diffuse’ regulation and also across the public sector. | 提言5:AI5原則が、特定の規制当局が存在しない領域や「拡散的」規制、また公共部門全体でどのように実施されるかを明らかにする。 |
Capability | 能力 |
Regulating AI is resource-intensive and highly technical. Regulators, civil society organisations and other actors need new capabilities to properly carry out their duties. | AIの規制は、資源集約的で高度な技術を要する。規制当局、市民社会組織、その他のアクターは、その責務を適切に遂行するために新たな能力を必要としている。 |
Scope and powers | 課題:範囲と権限 |
Regulator mandates and powers vary greatly, and many will be unable to force AI users and developers to comply with all the principles. | 規制当局の権限や職務権限は様々であり、その多くはAIの利用者や開発者にすべての原則の遵守を強制することはできない。 |
Recommendation 6: Introduce a statutory duty for legislators to have regard to the principles, including strict transparency and accountability obligations. | 提言6:厳格な透明性と説明責任の義務を含む、原則に配慮する立法者の法的義務を導入する。 |
Recommendation 7: Explore the introduction of a common set of powers for regulators and ex ante, developer-focused regulatory capability. | 提言7:規制当局に共通の権限と、開発者に焦点を当てた事前規制能力の導入を検討する。 |
Recommendation 8: Clarify the law around AI liability, to ensure that legal and financial liability for AI risk is distributed proportionately along AI value chains. | 提言8:AIリスクに対する法的・金銭的責任がAIバリューチェーンに沿って比例的に配分されるよう、AI責任に関する法律を明確化する。 |
Resourcing | 課題:人材確保 |
AI is increasingly a core part of our digital infrastructure, and regulators need significantly more resourcing to address it. | AIはますますデジタルインフラの中核を担うようになってきており、規制当局はAIに対応するために大幅な人員増を必要としている。 |
Recommendation 9: Significantly increase the amount of funding available to regulators for responding to AI-related harms, in line with other safety-case based regulatory domains. | 提言9:他のセーフティケースに基づく規制領域と同様に、AI関連の危害に対応するために規制当局が利用できる資金量を大幅に増やす。 |
The regulatory ecosystem | 課題:規制のエコシステム |
Other actors such as consumer groups, trade unions, charities and assurance providers will need to play a central role in holding AI accountable. | 消費者団体、労働組合、慈善団体、保証プロバイダなどの他のアクターは、AIに説明責任を果たさせる上で中心的な役割を果たす必要がある。 |
Recommendation 10: Create formal channels to allow civil society organisations, particularly those representing vulnerable groups, to meaningfully feed into future regulatory processes, the work of the Foundation Model Taskforce and the AI Safety Summit. | 提言10:市民社会組織、特に脆弱性グループの代表者が、将来の規制プロセス、財団モデル・タスクフォースの作業、AI安全サミットに有意義な情報を提供できるよう、正式なチャンネルを設ける。 |
Recommendation 11: Establish funds and pooled support to enable civil society organisations like consumer groups, trade unions and advisory organisations to hold those deploying and using AI accountable. | 提言11:消費者団体、労働組合、諮問機関のような市民社会組織が、AIを配備・使用 する側に説明責任を負わせることができるよう、資金とプールされた支援を確立する。 |
Recommendation 12: Support the development of non-regulatory tools such as standards and assurance. | 提言12:標準や保証などの非規制ツールの開発を支援する。 |
Urgency | 緊急性 |
The widespread availability of foundation models such as GPT-4 is accelerating AI adoption and risks scaling up existing harms. The Government, regulators and the Foundation Model Taskforce need to take urgent action. | GPT-4のような基盤モデルの普及は、AIの導入を加速させ、既存の被害を拡大させるリスクとなっている。ガバナンス、規制当局、ファウンデーションモデル・タスクフォースは、早急に行動を起こす必要がある。 |
Legislation and enforcement | 課題:法整備と施行 |
New legislation, and more robust enforcement of existing laws, will be necessary to ensure foundation models are safe. | 基盤モデルの安全性を確保するためには、新たな法整備と既存法のより強固な執行が必要である。 |
Recommendation 13: Allocate significant resource and future parliamentary time to enable a robust, legislatively supported approach to foundation model governance as soon as possible. | 提言13:財団モデルガバナンスに対する強固で、法制化されたアプローチを早急に実現するために、多大な資源と将来の国会の時間を割く。 |
Recommendation 14: Review opportunities for and barriers to the enforcement of existing laws – particularly the UK GDPR and the intellectual property (IP) regime – in relation to foundation models and applications built on top of them. | 提言14:ファウンデーションモデルとその上に構築されるアプリケーションに関連して、既存法(特に英国のGDPRと知的財産(IP)制度)の執行の機会と障壁を見直す。 |
Transparency and monitoring | 課題:透明性とモニタリング |
Too often, foundation models are opaque ‘black boxes’, with limited information available to the Government and regulators. | 多くの場合、基盤モデルは不透明な「ブラックボックス」であり、政府や規制当局が利用できる情報は限られている。 |
Recommendation 15: Invest in pilot projects to improve Government understanding of trends in AI research, development and deployment. | 提言15:AIの研究、開発、導入の動向に関する政府の理解を深めるため、パイロット・プロジェクトに投資する。 |
Recommendation 16: Introduce mandatory reporting requirements for developers of foundation models operating in the UK or selling to UK organisations. | 提言16:英国で活動する、あるいは英国の組織に販売する基盤モデルの開発者に対して、報告義務を導入する。 |
Leadership | 課題:リーダーシップ |
Priorities for AI development are currently set by a relatively small group of large industry players. | AI開発の優先順位は現在、比較的少数の業界大手グループによって決定されている。 |
Recommendation 17: Ensure the AI Safety Summit reflects diverse voices and an expansive definition of ‘AI safety’. | 提言17:AI安全性サミットが多様な声を反映し、「AI安全性」の定義が拡大されるようにする。 |
Recommendation 18: Consider public investment in, and development of, AI capabilities to steer applications towards generating long-term public benefit. | 提言18:長期的な公益を生み出す方向にアプリケーションを誘導するために、AI能力への公共投資とその開発を検討する。 |
・2023.07.18 Policy briefing: Regulating AI in the UK
P+B69:C80olicy briefing: Regulating AI in the UK | 政策ブリーフィング 英国におけるAI規制 |
This briefing examines the UK’s current plans for AI regulation an sets out recommendations for the Government and the Foundation Model Taskforce | 本ブリーフィングは、人工知能(AI)規制に関する英国の現行計画を検証し、政府と財団モデル・タスクフォースへの提言を示すものである。 |
This briefing examines the UK’s current plans for artificial intelligence (AI) regulation as set out in the March 2023 white paper ‘A pro-innovation approach to regulating AI’. It sets out 18 recommendations for the Government and the Foundation Model Taskforce that, if acted on, will help to strengthen the proposed regulatory framework. | 本ブリーフィングは、2023年3月の白書「A pro-innovation approach to regulating AI」で示された人工知能(AI)規制に関する英国の現行計画を検証する。本ブリーフィングは、政府とファウンデーション・モデル・タスクフォースに対する18の提言を示しており、これが実行されれば、提案されている規制の枠組みを強化するのに役立つだろう。 |
It is accompanied by a longer report – Regulating AI in the UK – which further contextualises and summarises the Government’s proposals.1 | 本報告書には、政府の提案の背景をさらに説明し要約した、より長い報告書「Regulating AI in the UK(英国におけるAIの規制)」も添付されている1。 |
The Ada Lovelace Institute is an independent research institute with a mission to make data and AI work for people and society. This means making sure that the opportunities, benefits and privileges generated by data and AI are justly and equitably distributed. | エイダ・ラブレス研究所は、データとAIを人々と社会のために役立てることを使命とする独立研究機関である。エイダ・ラブレス研究所は、データとAIを人々と社会のために機能させることを使命とする独立研究機関である。これは、データとAIが生み出す機会、利益、特権が公正かつ公平に分配されるようにすることを意味する。 |
If you would like more information on this policy briefing, or if you would like to discuss implementing our recommendations, please contact our policy research team at [mail]. | 本政策ブリーフィングに関する詳細情報をご希望の方、または我々の提言の実施について議論されたい方は、政策研究チーム[mail] までご連絡いただきたい。 |
・2023.07.18 AI risk: Ensuring effective assessment and mitigation across the AI lifecycle
AI risk: Ensuring effective assessment and mitigation across the AI lifecycle | AIリスク:AIのライフサイクルを通じた効果的な評価と低減の確保 |
Reducing the risks that AI systems pose to people and society | AIシステムが人と社会にもたらすリスクを軽減する |
Executive summary | 要旨 |
With the increasing use of AI systems in our everyday lives, it is essential to understand the risks they pose and take necessary steps to mitigate them. Because the risks of AI systems may become manifest at different stages of their deployment, and the specific kinds of risks that may emerge will depend on the contexts in which those systems are being built and deployed, assessing and mitigating risk is a challenging proposition. | 日常生活におけるAIシステムの利用が増加する中、AIシステムがもたらすリスクを理解し、それを低減するために必要な措置を講じることが不可欠である。AIシステムのリスクは、その展開のさまざまな段階で顕在化する可能性があり、顕在化する可能性のある具体的なリスクの種類は、システムが構築・展開される状況によって異なるため、リスクのアセスメントと低減は困難な命題である。 |
Addressing that challenge requires identifying and deploying a range of methods to be used by different actors across the lifecycle of an AI system’s development and deployment.1 By understanding these methods in more detail, policymakers and regulators can support their use by different actors in the UK’s technology sector, and so reduce the risks that AI systems can pose to people and society. | この課題に対処するには、AIシステムの開発と展開のライフサイクル全体にわたって、さまざまな関係者が使用するさまざまな手法を特定し、展開する必要がある1。これらの手法をより詳細に理解することで、政策立案者と規制当局は、英国のテクノロジー・セクターのさまざまな関係者によるこれらの手法の使用を支援し、AIシステムが人々と社会にもたらすリスクを低減することができる。 |
In its March 2023 AI regulation white paper, the UK Government proposed creating a set of central Government functions to support the work of regulators. This included a cross-sectoral risk-assessment function, intended to support regulators in their own risk assessments, to identify and prioritise new and emerging risks, and share risk enforcement best practices. This central function has the potential to help coordinate and standardise the somewhat fragmented risk-assessment landscape identified in this paper, and support the development of a cross-sectoral AI assessment ecosystem in the UK. | 英国ガバナンスは、2023年3月のAI規制白書で、規制当局の業務をサポートする一連の中央政府機能の創設を提案した。この機能には、規制当局のリスクアセスメントを支援し、新たなリスクを特定して優先順位を付け、リスク執行のベストプラクティスを共有することを目的とした、分野横断的なリスクアセスメント機能が含まれている。この中央機能は、本稿で明らかにされたやや断片的なリスクアセスメントの状況を調整・標準化し、英国における分野横断的なAIアセスメント・エコシステムの発展を支援する可能性を秘めている。 |
Key takeaways | 主な要点 |
1. There is not a singular, standardised process for assessing the risks or impacts of AI systems (or a common vocabulary), but there are commonly used components: policymakers, regulators and developers will need to consider how these are delivered and tailored. | 1. AIシステムのリスクや影響を評価するための唯一で標準化されたプロセス(または共通の語彙)は存在しないが、一般的に使用される構成要素は存在する:政策立案者、規制当局、開発者は、これらをどのように提供し、調整するかを検討する必要がある。 |
2. Risk and impact assessment methods typically involve five components: risk identification, risk prioritisation, risk mitigation planning, risk monitoring and communicating risks. The main differences between components are in how they are achieved, the actors involved, the scope of impacts considered and the extent of accountability. | 2. リスクと影響のアセスメント手法には、通常5つの要素が含まれる: リスクの特定、リスクの優先順位付け、リスク低減計画、リスクのモニタリング、リスクのコミュニケーションである。構成要素間の主な違いは、その達成方法、関与する主体、考慮する影響の範囲、説明責任の範囲にある。 |
3. Policymakers globally are incorporating risk and impact assessments in AI governance regimes and legislation, with the EU, USA, Brazil and Canada mandating assessments for various AI systems. Regulators and policymakers face the challenge of ensuring risk consideration is conducted, acted on and monitored over time, highlighting the need for an ecosystem of assessment methods. | 3. 世界の政策立案者は、AIのガバナンス体制や法律にリスクアセスメントやインパクトアセスメントを取り入れている、 EU、米国、ブラジル、カナダでは、様々なAIシステムに対して評価を義務付けている。規制当局や政策立案者は、リスクアセスメントが確実に実施され、実施され、長期にわたってモニタリングされるという課題に直面しており、アセスメント手法のエコシステムの必要性が浮き彫りになっている。 |
4. Identifying and assessing risks alone does not ensure risks are avoided. AI risk management will require an ecosystem of assessment, assurance and audit, including independent auditing, oversight bodies, ethics review committees, safety checklists, model cards, datasheets and transparency registers, which collectively enable monitoring and mitigation of AI-related risks. | 4. リスクの識別とアセスメントだけでは、リスクを確実に回避することはできない。 AIのリスクマネジメントには、独立監査、監視団体、倫理審査委員会、安全性チェックリスト、モデルカード、データシート、透明性登録など、評価、保証、監査のエコシステムが必要であり、これらは総体としてAI関連リスクの監視と低減を可能にする。 |
5. Ensuring this AI assessment ecosystem is effective will require consensus on risk-assessment standards, supported by incentives for assessing societal risks and case studies showcasing risk-assessment methods in practice. Domain-specific guidance, skilled professionals and strong regulatory capacity can further enhance the ecosystem. Third-party assessors, including civil society, academia and commercial services, will be essential for developing and implementing assessment practices at scale. | 5. このAIアセスメント・エコシステムが効果的であることを保証するには、リスクアセスメント標準に関するコンセンサスが必要であり、社会的リスクを評価するためのインセンティブや、リスクアセスメント手法の実践事例を紹介するケーススタディに支えられる必要がある。分野に特化したガイダンス、熟練した専門家、強力な規制能力は、エコシステムをさらに強化することができる。市民社会、学術界、商業サービスを含むサードパーティ・アセッサーは、アセスメント手法を開発し、大規模に実施するために不可欠である。 |
Research questions | 研究課題 |
・What are the broad areas of risks that AI systems can pose in different contexts (particularly from emerging AI technologies)? | ・AIシステム(特に新興AI技術)がさまざまな文脈でもたらしうるリスクには、どのようなものがあるか? |
・How should regulators or policymakers respond to different kinds of risks? | ・規制当局や政策立案者は、さまざまな種類のリスクにどのように対応すべきか。 |
・What mechanisms and processes can be used to assess different kinds of risks, including the significance of their potential impact and their likelihood? | ・潜在的な影響や可能性の重要性を含め、さまざまな種類のリスクを評価するために、どのようなメカニズムやプロセスを用いることができるのか。 |
・Whose responsibility (e.g. developer, procurer, regulator) is it to conduct these assessments and evaluations? | ・これらの評価や査定は誰の責任(開発者、調達者、規制当局など)で行うのか? |
・What are methods for checking, monitoring and mitigating risks through the AI lifecycle? | ・AIのライフサイクルを通じてリスクをチェックし、監視し、低減する方法にはどのようなものがあるか? |
・What might be needed for an effective assessment ecosystem? | ・効果的な評価エコシステムには何が必要だろうか? |
To answer these questions, this paper surveys approaches for assessing risks that AI systems pose for people and society – both on the ground within AI project teams, and in emerging legislation. The findings of this report are based on a desk-based review and synthesis of grey and academic literature on approaches to assessing AI risk, alongside analysis of draft regulations that contain requirements for anticipating risk or impacts of AI systems. | これらの問いに答えるため、本稿では、AIプロジェクトチームの現場と、新たな法規制の両方において、AIシステムが人々や社会にもたらすリスクをアセスメントするためのアプローチを調査する。本報告書の調査結果は、AIシステムのリスクや影響を予測するための要件を含むドラフト規制の分析とともに、AIリスク評価のアプローチに関する灰色文献および学術文献の机上レビューと統合に基づいている。 |
Key terms | 主な用語 |
Impact assessment: Impact assessments are evaluations of an AI system that use prompts, workshops, documents and discussions with the developers of an AI system and other stakeholders to explore how a particular AI system will affect people or society in positive or negative ways. These tend to occur in the early stages of a system’s development before it is in use, but may occur after a system has been deployed. | 影響評価: 影響評価とは、AIシステムの評価であり、プロンプト、ワークショップ、文書、AIシステムの開発者やその他の利害関係者とのディスカッションなどを用いて、特定のAIシステムが人々や社会にどのようなプラスまたはマイナスの影響を与えるかを探るものである。これらは、システムが使用される前の開発の初期段階で行われることが多いが、システムが導入された後に行われることもある。 |
Risk assessment: Risk assessments are very similar to impact assessments, but look specifically at the likelihood of harmful outcomes occurring from an AI system. These also tend to occur in the early stages of a system’s development before it is in use, but may occur after a system has been deployed. | リスクアセスメント: リスクアセスメントはインパクトアセスメントとよく似ているが、特にAIシステムによって有害な結果が生じる可能性を検討するものである。これも、使用前のシステム開発の初期段階で行われる傾向があるが、システムが配備された後に行われることもある。 |
Algorithmic audit: Algorithmic audits are a form of external scrutiny of an AI system, or the processes around it, which can be conducted as part of a risk or impact assessment. These can be technical audits of the inputs or outputs of a system; compliance audits of whether an AI development team has completed processes or regulatory requirements; regulatory inspections by regulators to monitor behaviour of an AI system over time; or sociotechnical audits that evaluate the ways in which a system is impacting wider societal processes and contexts in which it is operating. Audits usually occur after a system is in use, so can serve as accountability mechanisms to verify whether a system behaves as developers intend or claim. | アルゴリズム監査: アルゴリズム監査は、リスクアセスメントや影響度アセスメントの一環として実施される、AIシステムまたはその周辺のプロセスに対する外部からの精査の一形態である。システムのインプットやアウトプットの技術的な監査、AI開発チームがプロセスや規制要件を完了しているかどうかのコンプライアンス監査、AIシステムの挙動を長期的に監視するための規制当局による規制検査、あるいはシステムが動作しているより広範な社会のプロセスやコンテクストにどのような影響を与えているかを評価する社会技術的な監査などがある。監査は通常、システムが使用された後に行われるため、システムが開発者の意図や主張通りに動作するかどうかを検証する説明責任のメカニズムとして機能する。 |
・2023.07.18 Keeping an eye on AI
Keeping an eye on AI | AIから目を離さない |
Approaches to government monitoring of the AI landscape | 政府によるAI監視のガバナンス |
Executive summary | 要旨 |
The rapid development and deployment of artificial intelligence (AI) systems has the potential to be transformative for society. Whether its effects are beneficial or harmful, AI comes with an array of challenges that policymakers must navigate. To stay ahead of the curve and make well informed strategic decisions, it is essential that the UK Government possesses accurate and timely information about AI systems and their applications. | 人工知能(AI)システムの急速な開発と普及は、社会に変革をもたらす可能性を秘めている。その効果が有益なものであれ有害なものであれ、AIには政策立案者が乗り越えなければならない様々な課題が伴う。曲線の先端に立ち、十分な情報に基づいた戦略的決定を行うためには、英国政府がAIシステムとその応用に関する正確でタイムリーな情報を保有することが不可欠である。 |
Just as governments closely monitor other sectors to inform policy, a similar approach can be adopted for AI. Consider how governments use inflation statistics to guide economic policymaking, or the creation of COVID-19 dashboards for public health decisions. These examples showcase the ability to aggregate complex data and distil it into actionable insights, which can then be used to shape effective policies. | 政府が政策に反映させるために他のセクターを注意深く監視しているように、同様のアプローチをAIにも採用することができる。政府が経済政策立案のためにインフレ統計をどのように利用しているか、あるいは公衆衛生の意思決定のためにCOVID-19ダッシュボードをどのように作成しているかを考えてみよう。これらの例は、複雑なデータを集約し、実行可能な洞察に絞り込む能力を示している。 |
In the context of AI, effective monitoring could provide crucial insights into the impacts, risks and opportunities of AI systems across different sectors and countries. By examining AI’s influence on areas such as employment and recruitment, finance, healthcare and critical infrastructure, the Government could make decisions that maximise benefits and mitigate potential injustices or harms. For instance, monitoring could reveal the need for new regulations to protect affected persons, or investments in safety research for foundation models. | AIの文脈では、効果的なモニタリングによって、さまざまな分野や国にわたるAIシステムの影響、リスク、機会について重要な洞察を得ることができる。雇用・採用、金融、医療、重要インフラなどの分野におけるAIの影響力を調査することで、政府は利益を最大化し、潜在的な不正や害を軽減する決定を下すことができる。例えば、モニタリングによって、影響を受ける人々を保護するための新たな規制や、基盤モデルの安全性研究への投資の必要性が明らかになるかもしれない。 |
Despite the clear need for comprehensive AI monitoring capabilities, the Government currently lacks the necessary infrastructure and tools to effectively track AI developments. By investing in robust AI monitoring systems, both the Government and the public will be better equipped to make informed decisions, based on high-quality data, about whether to use or trust AI systems. This would create a deeper understanding of the societal impact of the use of AI systems and create a robust base for forward-thinking policies. | 包括的なAIモニタリング機能の必要性は明らかであるにもかかわらず、政府は現在、AIの発展を効果的に追跡するために必要なインフラやツールを欠いている。強固なAIモニタリング・システムに投資することで、政府も国民も、AIシステムを利用するか信頼するかについて、質の高いデータに基づき、十分な情報に基づいた意思決定を行うことができるようになる。これにより、AIシステムの利用が社会に与える影響についてより深い理解が得られ、将来を見据えた政策のための強固な基盤が構築されるだろう。 |
The key question of this paper is: | 本稿の重要な問いは以下の通りである: |
Given the speed of progress and complexity in AI development and deployment, how can Government source, rapidly synthesise and summarise information about technological capabilities, trends, risks and opportunities in order to make informed, strategic policy decisions? | AIの開発・導入における進歩の速さと複雑さを考えると、政府はどのようにして、技術的能力、トレンド、リスク、機会に関する情報を入手し、迅速に統合し、要約して、情報に基づいた戦略的な政策決定を行うことができるのか? |
Key takeaways | 主な要点 |
1. There are specific properties of AI systems that the Government should consider measuring and monitoring such as their deployment and governance, and their downstream impacts on individuals and society. This information could include: inputs to AI systems; categorical information about the data and model underlying the AI systems; categorical information about processes or operations followed in development and deployment; direct outputs and outcomes of AI systems; and externalities generated by those outputs and outcomes. | 1. AIシステムには、その展開やガバナンス、個人や社会に対する下流への影響など、政府が測定・監視を検討すべき特定の特性がある。 この情報には、AIシステムへのインプット、AIシステムの基礎となるデータやモデルに関するカテゴリー情報、開発・展開のプロセスやオペレーションに関するカテゴリー情報、AIシステムの直接的なアウトプットや成果、アウトプットや成果によって生み出される外部性が含まれる。 |
2. The Government is well placed to address gaps in the existing monitoring ecosystem. This could be through standardised and mandated disclosure of information from companies, g. on compute; voluntary or statutory sharing of commercially sensitive information; and working with other governments on global comparative monitoring efforts. However, not all this information could or should be collected directly by central Government. Sectoral and cross-cutting regulators will have an essential role in providing contextual information during the gathering and interpretation of quantitative or aggregate data on AI capabilities and risks. | 2. 政府は、既存のモニタリング・エコシステムにおけるギャップに対処することができる。これは、企業からの情報開示の標準化・義務化(計算機など)、商業上機微な情報の自主的・法定的共有、世界的な比較モニタリングの取り組みにおける他国政府との協力などを通じて行うことができる。しかし、こうした情報のすべてを中央政府が直接収集できるわけではないし、収集すべきでもない。セクターや分野横断的な規制当局は、AIの能力やリスクに関する定量的・集計的データの収集や解釈の際に、文脈に沿った情報をプロバイダとして提供する上で不可欠な役割を担うことになる。 |
3. Regulators already have a number of existing mechanisms to ensure that information about the AI landscape is identified and shared with them. This information includes new developments, opportunities and risks with a sector or domain. Possible mechanisms are: standards for model cards and datasheets; regulatory sandboxes or multi-agency advisory services, regulatory inspection and audit; whistleblower protections and rewards; incident reporting; and ombudsmen. | 3. 規制当局には、AIの状況に関する情報が特定され、共有されるようにするための既存のメカニズムが既にいくつかある。この情報には、分野や領域に関する新たな進展、機会、リスクなどが含まれる。考えられるメカニズムとしては、モデルカードやデータシートの標準、規制当局のサンドボックスや複数機関による助言サービス、規制当局の検査や監査、内部告発者の保護と報奨、インシデント報告、オンブズマンなどがある。 |
4. Foundation models present unique challenges for cross-sectoral regulation. If and when AI applications in different sectors become more reliant on foundation models, it will not be efficient for individual regulators to each individually assess and monitor these systems and create multiple, overlapping and potentially conflicting demands on the companies deploying them. One option to address this issue is the creation of a centralised AI regulatory function with institutional relationships with the relevant sectoral or cross-cutting regulators and direct monitoring relationships with developers of foundation models. | 4. ファウンデーション・モデルは、分野横断的な規制のためのユニークな課題を提示する。 異なるセクターにおけるAIの応用が基盤モデルへの依存度を高めた場合、個々の規制当局がそれぞれ個別にこれらのシステムを評価・監視することは効率的ではなく、導入企業に対して複数の、重複した、そして潜在的に矛盾する要求を生み出すことになる。この問題に対処するための一つの選択肢は、関連するセクターや分野横断的な規制当局と機構的な関係を持ち、基盤モデルの開発者と直接的な監視関係を持つ、一元化されたAI規制機能の創設である。 |
5. The UK Government should not delay in building out its internal monitoring capabilities, and should immediately initiate small, focused pilot projects responding to policy challenges. This could include establishing a national-level public repository of the harms, failures and unintended negative consequences of AI systems; building on the Review of the Future of Compute; putting in place Government monitoring, aggregating (and potentially publication) of data on broad compute use and demand trends; and requesting to be informed when frontier AI labs begin large-scale training runs of new models. | 5. 英国政府は、内部監視能力の構築を遅らせてはならず、政策課題に対応する小規模で集中的なパイロット・プロジェクトを直ちに開始すべきである。 これには、AIシステムの危害、失敗、予期せぬ悪影響に関する国家レベルの公的レポジトリの設立、「コンピ ューターの将来に関するレビュー」の構築、幅広いコンピュート使用と需要の傾向に関するデータの政府による監視、集計(可能であれば 公表)、フロンティアAIラボが新モデルの大規模な訓練を開始する際の情報提供の要請などが含まれる。 |
Research questions | 研究課題 |
This paper aims to understand how Government can better measure and monitor developments in the AI landscape, addressing the following research questions: | 本稿の目的は、政府がAIの動向をより適切に測定・監視する方法を理解することであり、以下の研究課題に取り組むことである: |
・What value can monitoring AI developments provide? | ・AI開発のモニタリングはどのような価値をプロバイダに提供できるか? |
・What aspects of AI research, development and deployment could be measured? | ・AIの研究、開発、展開のどのような側面を測定することができるか? |
・How could the Government monitor the AI landscape? | ・政府はどのようにAIの状況を監視できるか? |
・What are the monitoring gaps in the existing ecosystem that the Government is well placed to fill? | ・既存のエコシステムにおけるモニタリングのギャップのうち、政府が埋められるものは何か? |
・What mechanisms are there for sectoral and cross-cutting monitoring by individual regulators? | ・個々の規制当局による分野別・横断的な監視にはどのような仕組みがあるか? |
・How can the Government address the challenges of monitoring across development and deployment of AI foundation models? | ・政府は、AI基盤モデルの開発・展開を横断的にモニタリングするという課題にどのように対処できるのか? |
This paper provides an analysis of approaches to monitoring developments in the AI landscape and outlines elements of AI systems that the Government could monitor. | 本稿では、AIの発展を監視するアプローチの分析を行い、政府が監視できるAIシステムの要素を概説する。 |
It then analyses existing private and intergovernmental initiatives for systematically monitoring AI development and deployment, and identifies several gaps that the Government is uniquely be able to fill. It then examines approaches to monitoring and responsible disclosure in individual sectors, including sandboxes, incident reporting and horizon scanning. Finally, this paper concludes with a discussion of how these approaches could be complicated by the introduction of general-purpose AI systems and future questions that need to be explored. | 次に、AIの開発と展開を体系的にモニタリングするための既存の民間および政府間のイニシアティブを分析し、政府が独自に埋めることのできるいくつかのギャップを特定する。続いて、サンドボックス、インシデント報告、ホライズン・スキャンなど、各分野におけるモニタリングと責任ある情報開示へのアプローチを検証する。最後に、これらのアプローチが汎用AIシステムの序文によってどのように複雑化しうるか、また今後検討すべき課題について考察し、本稿を締めくくる。 |
Key terms | 主要用語 |
Measurement is the collection of information that reduces (expected) uncertainty about a given topic. This requires deciding what information needs to be collected and how to best gather or collect it.1 | 測定とは、与えられたトピックに関する(予想される)不確実性を低減する情報の収集である。そのためには、どのような情報を収集する必要があるのか、また、どのように収集するのが最適なのかを決定する必要がある1。 |
Monitoring is the process of operationalising measurement over time and entails systemically and continually gathering measurements in a common format.2 Monitoring allows for tracking changes over time and for that information to be aggregated and integrated in policymaking. | モニタリングは、測定値を長期的に運用するプロセスであり、体系的かつ継続的に共通のフォーマットで測定値を収集することを必要とする2。モニタリングによって経年変化を追跡し、その情報を集約して政策立案に反映させることができる。 |
Ex ante mechanisms are forward-looking regulatory tools that take effect before an AI system is deployed and impacts users and affected people. Examples of ex-ante mechanisms include regulatory sandboxes which allow companies and regulators to test AI products in a real-world environment before they enter the market. | 事前的メカニズムとは、AIシステムが導入され、利用者や影響を受ける人々に影響を与える前に効力を発揮する、将来を見据えた規制手段である。事前規制の例としては、企業や規制当局がAI製品を市場に投入する前に現実の環境でテストすることを可能にする規制のサンドボックスなどがある。 |
Ex post mechanisms are backwards-looking regulatory tools that take effect after an AI system is deployed. These include regulatory inspection and auditing methods, in which an AI system is evaluated and tested by a regulator for compliance with a particular law or regulation. | 事後的メカニズムとは、AIシステムが導入された後に効力を発揮する、後ろ向きの規制ツールである。これには、AIシステムが特定の法律や規制を遵守しているかどうかを規制当局が評価・テストする、規制当局の検査・監査手法が含まれる。 |
● まるちゃんの情報セキュリティ気まぐれ日記
英国のAI規制...
・2023.07.20 英国 Ada Lovelace 協会 英国のAI規制、AIリスクなど4つの報告書
・2023.07.18 英国 科学技術省データ倫理・イノベーションセンターブログ AI保証の専門家が他の領域から学ぶべき6つの教訓 (2023.07.12)
・2023.07.09 英国 著作権と人工知能 (AI) に関する実施規範をつくるようですね。。。(2023.06.29)
・2023.07.07 英国 Ada Lovelace 協会 AIサプライチェーンにおける説明責任の分担 (2023.06.29)
・2023.06.19 英国 科学技術省データ倫理・イノベーションセンター 「AIシステムをより公平にするために、人口統計データへの責任あるアクセスを可能にする」
・2023.06.18 英国 科学技術省データ倫理・イノベーションセンター AI保証事例集 (2023.06.07)
・2023.04.29 英国 AIに関する英国政府の発表をいくつか。。。
・2023.04.01 英国 意見募集 AI規制白書
・2023.03.10 英国 ICO Blog 国際女性デーに向けて...AIによる差別への対処が重要な理由
・2022.12.10 英国 データ倫理・イノベーションセンター「業界温度チェック:AI保証の障壁と実現要因」
・2022.11.06 英国 データ倫理・イノベーションセンター データおよびAIに対する国民の意識:トラッカー調査(第2回)
・2022.10.29 英国 ICO 雇用慣行とデータ保護:労働者の健康情報案に関するコンサルテーション (2022.10.27)
・2022.09.25 英国 Ada Lovelace 協会 欧州におけるAI責任:EUのAI責任指令の先取り
・2022.06.18 英国 Ada Lovelace 協会: EUのAI法について説明 (2022.04.11)
・2021.06.22 人工知能技術に芸術と人文科学はどのように貢献できるか? by John Tasioulas(オックスフォード大学のAI倫理研究所所長)
・2020.05.01 (人工知能 AI)ブラックボックスの検証:アルゴリズムシステムを評価するためのツール - アルゴリズムの監査・影響評価のための共通言語の特定
Comments