欧州議会 Think Tank 欧州連合(EU)のAI行動規範 (2025.08.27)
こんにちは、丸山満彦です。
AI行動規範についての欧州議会のThink Tankからの説明...
● European Parliament - Think Tank
・2025.08.27 The European Union's AI code of practice
| The European Union's AI code of practice | 欧州連合(EU)のAI行動規範 |
| As part of its efforts to regulate artificial intelligence (AI), the EU recently developed a code of practice for general-purpose AI (GPAI) models. This code details how to implement the obligations related to GPAIs as set out in the AI Act. Being key to the AI Act's implementation, the code has been at the centre of heated discussions. | 人工知能(AI)規制の一環として、EUは汎用AI(GPAI)モデル向けの行動規範を策定した。この規範は、AI法で定められたGPAI関連の義務をどのように履行するかを詳細に規定している。AI法実施の要となる本規範は、激しい議論の的となっている。 |
| Regulating GPAI models | GPAIモデルの規制 |
| Adopted in 2024, the AI Act aims to improve the functioning of the internal market, promote a humancentric approach to AI, limit risks and support innovation. The AI Act distinguishes between AI systems and GPAI models and lays out a common framework for their use and supply in the EU. AI systems are machinebased systems capable of inferring an output based on an input, while GPAI models are AI models trained on a large amount of data and able to perform a wide range of distinct tasks. Moreover, GPAI models may present a systemic risk when they meet the conditions specified in Article 51 of the AI Act. The AI Act lists additional obligations for GPAI with systemic risks, such as the obligations to perform model evaluation, assess and mitigate possible systemic risks, and ensure cybersecurity. | 2024年に採択されたAI法は、域内市場の機能改善、人間中心のAIアプローチの推進、リスクの抑制、イノベーション支援を目的としている。AI法はAIシステムとGPAIモデルを区別し、EU域内での使用・供給に関する共通枠組みを定めている。AIシステムとは入力に基づいて出力を推論できる機械ベースのシステムを指す一方、GPAIモデルは大量のデータで訓練され多様なタスクを実行可能なAIモデルである。さらにGPAIモデルは、AI法第51条に規定された条件を満たす場合、システミックリスクをもたらす可能性がある。AI法は、システムリスクを伴うGPAIに対して、モデル評価の実施、潜在的なシステムリスクの、緩和、サイバーセキュリティの確保といった追加義務を定めている。 |
| Most of the AI Act's provisions will come into effect several months or years after its adoption. The AI Act also establishes several deadlines for instruments facilitating its implementation, with the code of practice being one of these instruments. The code specifically covers GPAI obligations (see section below), and the European Commission has chosen independent experts to draft it. The AI Act states that the code was to be ready by 2 May 2025 at the latest, but this deadline has passed without adoption. | AI法の規定の大部分は、採択後数か月から数年を経て AI法はまた、その実施を促進する手段(行動規範もその一つ)について複数の期限を設定している。行動規範は特にGPAIの義務(下記セクション参照)を規定しており、欧州委員会は独立専門家を選定してそのドラフトを委託した。AI法では行動規範は遅くとも2025年5月2日までに完成させることとしていたが、この期限は過ぎても採択されていない。 |
| A code of practice to implement GPAI obligations | GPAI義務を実施するための行動規範 |
| The AI Act establishes high-level legal obligations related to all types of GPAI models, such as adopting policies to ensure compliance with EU law on copyright; the relevant stakeholders may use different technical means to meet these obligations. The code of practice is a non-binding document that provides guidance on how providers can technically implement their GPAI obligations. | AI法は、著作権に関するEU法の遵守を確保するためのポリシー採用など、あらゆる種類のGPAIモデルに関連する高水準の法的義務を定めている。関係するステークホルダーは、これらの義務を満たすために異なる技術的手段を用いることができる。行動規範は、プロバイダがGPAI義務を技術的にどのように実施できるかについてのガイダンスを提供する、拘束力のない文書である。 |
| Article 56 of the AI Act states that the AI Office of the Commission and the AI Board (the EU-level governance body set up by the AI Act) must ensure that the code of practice covers at least the obligations applying to GPAI models, including those presenting systemic risks. Article 56 adds that the AI Office may encourage the participation of relevant stakeholders in the drafting process of the code. However, this article says that GPAI providers may be invited to adhere to the code of practice. | AI法第56条は、欧州委員会のAI事務局およびAI委員会(AI法により設置されたEUレベルのガバナンス機関)が、行動規範が少なくともGPAIモデル(システミックリスクをもたらすものを含む)に適用される義務を網羅するよう確保しなければならないと規定している。第56条はさらに、AI事務局が行動規範のドラフトプロセスへの関係者の参加を促すことができると付記している。ただし、同条はGPAIプロバイダが行動規範の遵守を要請される可能性があるとしている。 |
| The challenge of operationalising the rules | 規則の実践化における課題 |
| According to some of the experts involved in drafting the code, 'it is the first time that legal rules are turned into more detailed guidelines for the responsible development and deployment of GPAI'. However, defining the requirements and practical measures is challenging, because GPAI providers, businesses and civil society representatives have differing preferences regarding the solutions that the code should promote. | 規範ドラフトに関与した一部の専門家によれば、「汎用AIの責任ある開発・展開に向けた法的規則が詳細なガイドラインに転換されるのは今回が初めて」である。しかし、汎用AIプロバイダ、企業、市民社会代表者が規範が推進すべき解決策について異なる選好を持つため、要件と実践的措置の定義は困難を伴う。 |
| The AI Act obligation for GPAI providers to put in place a policy to comply with EU copyright law and related rights is a clear example of differing preferences. To contribute to its proper application, the AI Act mandates the code to address this obligation. On the one hand, rights-holders would prefer the code to implement strict technical measures on the use of copyrighted works for GPAI training. On the other hand, GPAI providers would prefer measures providing them with more freedom. Due to legal uncertainties, both approaches could potentially be incorporated into the code. | EU著作権法及び関連権利への準拠方針策定義務(AI法規定)は、こうした相違の顕著な例である。同法の適切な適用を促進するため、AI Actはプロバイダに対し、この義務への対応を定める行動規範の制定を義務付けている。権利者側は、GPAI訓練における著作権作品利用に対し厳格な技術的措置を規範に盛り込むことを望む。一方、GPAIプロバイダはより自由度が高い措置を望んでいる。法的な不確実性のため、両方のアプローチが規範に組み込まれる可能性がある。 |
| The fourth and final version of the code includes 40 measures categorised under transparency, copyright, security and safety, with each measure consisting of one or more points. The code sets out five measures for GPAI providers to comply with EU copyright law. One measure is to 'identify and comply with rights reservations when crawling the World Wide Web', which includes five points. The first point specifies that signatories will 'employ web-crawlers that read and follow instructions expressed in accordance with the Robot Exclusion Protocol (robots.txt) ...' and 'identify and comply with other appropriate machine-readable protocols to express rights reservation...'. The use of robots.txt to express rights reservations has been controversial as it was not originally designed for this purpose. Rights-holders argue that this technology does not offer them enough possibilities. However, given their varying interests, rights-holders have not reached a consensus on a single solution to replace robots.txt. A researcher listed six other available standards, including the spawning protocol, HTTP response headers and Hypertext Markup Language. | 規範の第四版(最終版)には、透明性、著作権、セキュリティ・安全性の4分野に分類された40の措置が含まれており、各措置は1つ以上の項目で構成される。規範はGPAIプロバイダがEU著作権法を遵守するための5つの措置を定めている。その一つが「ワールドワイドウェブのクロール時に権利留保を識別し遵守する」という措置で、5つの項目を含む。最初の項目では、署名者が「ロボット排除プロトコル(robots.txt)に従って表現された指示を読み取り従うウェブクローラーを採用する...」こと、および「権利留保を表現するためのその他の適切な機械可読プロトコルを識別し遵守する...」ことを規定している。権利留保の表明にrobots.txtを使用することは、当初この目的のために設計されたものではないため議論の的となっている。権利保有者は、この技術が十分な可能性を提供していないと主張する。しかし、権利保有者の利害が異なるため、robots.txtに代わる単一の解決策について合意には至っていない。ある研究者は、スポーニングプロトコル、HTTPレスポンスヘッダー、ハイパーテキストマークアップ言語(HTML)など、他に利用可能な6つの標準を列挙した。 |
| The challenge of the new political context | 新たな政治的文脈における課題 |
| EU AI policies have evolved since the adoption of the AI Act, with a clear objective to make Europe a leading AI continent. In April 2025, the Commission released an AI continent action plan outlining a comprehensive strategy to achieve this goal. The action plan includes measures such as cutting red tape and simplifying EU rules for citizens and businesses. It also echoes one of the recommendations made by former Italian Prime Minister Mario Draghi in his 2024 report, aimed at closing the innovation gap with competitors. It forms part of a broader initiative – the competitiveness compass – which aims to increase the EU's competitiveness. | EUのAI政策はAI法の採択以来進化を続けており、欧州をAI分野の主導的大陸とする明確な目標を掲げている。2025年4月、欧州委員会はこの目標達成に向けた包括的戦略をまとめた「AI大陸行動計画」を発表した。この行動計画には、規制緩和や市民・企業向けのEU規則簡素化などの措置が含まれる。また、2024年の報告書でマリオ・ドラギ元イタリア首相が提言した、競合国とのイノベーション格差解消に向けた勧告の一つにも呼応している。これはEUの競争力向上を目指す広範なイニシアチブ「競争力コンパス」の一環を成す。 |
| As a result, the code of practice faces a challenging political situation. It is supposed to clarify rules from a regulation adopted during the Commission's previous mandate, yet the EU's priorities have shifted. There is tension between turning established legal rules into technical rules and, at the same time, trying to simplify them. Furthermore, several EU industries have called for a pause in the AI Act's implementation. The Commission answered in July 2025, specifying that there is no pause for the AI Act's legal deadlines, although the code of practice's implementation date can be discussed. The Commission suggested the end of 2025. | その結果、行動規範は困難な政治的状況に直面している。これは前委員会の任期中に採択された規制のルールを明確化するはずだったが、EUの優先事項は変化した。確立された法的ルールを技術的ルールに変換すると同時に、それらを簡素化しようとする試みとの間に緊張が生じている。さらに、複数のEU産業がAI法の実施の一時停止を求めている。欧州委員会は2025年7月に回答し、行動規範の実施時期については協議の余地があるものの、AI法の法的期限については停止はないと明言した。委員会は2025年末を提案した。 |
| The fourth and final version of the code of practice | 行動規範の第四版(最終版) |
| The independent experts started drafting the code in July 2024, basing their approach on stakeholder feedback. The Commission reported that nearly 1 400 participants took part in the first round of drafting. The code has evolved significantly throughout its four drafts – the three first drafts failed to satisfy stakeholders; the fourth and final version has received mixed reactions. Some representatives of GPAI providers and other large technology companies have criticised the code for introducing measures they consider beyond the scope of the AI Act. However, other representatives of large technology companies have welcomed the changes, seeing them as improvements. In a 16 July 2025 press release, the European Parliament’s AI Act working group co-chairs welcomed the finalisation of the code. Representatives of civil society, including rights-holders, have acknowledged the improvements but expressed dissatisfaction. | 独立専門家グループは2024年7月に規範のドラフトを開始し、そのアプローチはステークホルダーからのフィードバックに基づいた。欧州委員会は、第1次ドラフト作成に約1,400名の参加者が関与したと報告した。行動規範は4回のドラフト作成過程で大幅に変更された——最初の3ドラフトは関係者の要求を満たせず、第4版(最終版)には賛否両論の反応が寄せられている。GPAIプロバイダやその他大手テクノロジー企業の代表者らは、AI法の範囲を超えると考える措置を導入したとして行動規範を批判している。しかし、他の大手テクノロジー企業の代表者らは、変更点を改善と捉え歓迎している。2025年7月16日のプレスリリースで、欧州議会のAI法作業部会共同議長は規範の最終化を歓迎した。市民社会(権利者を含む)の代表者らは改善点を認めつつも不満を表明している。 |
| Next steps and considerations for the European Parliament | 欧州議会の今後の対応と検討事項 |
| The Members of the European Parliament are closely monitoring the AI Act's implementation, including the development of the GPAI code of practice, through a joint IMCO-LIBE working group. According to the cochairs, the group has also discussed interplay with other pieces of EU legislation and upcoming guidelines. | 欧州議会議員は、IMCO-LIB合同E 作業部会を通じて、GPAI行動規範の策定を含むAI法の実施状況を注視している。共同議長によれば、同部会では他のEU法規や今後のガイドラインとの相互関係についても議論が行われた。 |
| On 1 August 2025, the AI Office and the Board – composed of one representative per Member State – published their assessment of the adequacy of the final code of practice. The AI Office invited GPAI providers to adhere to the code and published the list of its signatories. Not all major GPAI providers have fully signed the code of practice yet. Meta – the provider of the Meta AI and Llama models – has refused to sign. However, securing the signatures of all major providers would be an important step in implementing the AI Act. Experts believe that a successful code of practice could 'set the standard for future AI safety rules worldwide', and that 'guaranteeing the safety of frontier AI is a shared goal by all key global leaders'. | 2025年8月1日、AI事務局と加盟国の代表者1名で構成される理事会は、最終版行動規範の妥当性に関するアセスメントを公表した。AI事務局はGPAIプロバイダーに対し規範遵守を要請し、署名者リストを公開した。主要GPAIプロバイダの全てが規範に完全署名したわけではない。Meta AIおよびLlamaモデルを提供するMetaは署名を拒否している。しかしながら、主要プロバイダ全員の署名確保はAI法の実施における重要な一歩となる。専門家らは、成功した行動規範が「将来の全世界的なAI安全標準の礎となる」と確信し、「フロンティアAIの安全確保は全ての主要グローバルリーダーの共通目標である」と指摘している。 |
| The Commission may approve the code of practice through an implementing act, granting it general validity in the EU. This procedure involves a committee of experts appointed by the Member States, who can reject the Commission's draft implementing act with a qualified majority vote. The AI Office may also facilitate the review of the code of practice, in the light of emerging standards. | 欧州委員会は実施法令を通じ行動規範を承認し、EU域内での普遍的妥当性を確認する可能性がある。この手続きには加盟国が任命する専門家委員会が関与し、委員会は過半数の賛成で欧州委員会の実施法令ドラフトを否決できる。AI事務局は新たな標準を踏まえた行動規範の見直しも促進する可能性がある。 |
| In a few years, the code of practice will be supplemented by the harmonised standards, which will bring further concrete guidance on how GPAI providers can comply with the AI Act. | 数年後には、行動規範は調和された標準によって補完され、GPAIプロバイダがAI法に準拠する方法についてさらなる具体的な指針が示される見込みである。 |
⚫︎ まるちゃんの情報セキュリティ気まぐれ日記
・2025.07.31 欧州委員会 AI法における汎用AIモデル提供者の義務範囲に関するガイドライン(2025.07.18)
・2025.07.25 欧州委員会 汎用 AI 実践規範 (2025.07.10)
・2025.04.11 欧州委員会 AI大陸行動計画 (2025.04.09)
・2025.03.13 欧州委員会 AI法における汎用AIモデル - Q&A (2025.03.10)
・2025.02.08 欧州委員会 規則(EU)2024/1689(AI法)が定める人工知能の禁止行為に関する欧州委員会ガイドライン
・2024.12.22 欧州委員会 汎用AI実践規範の第2ドラフトを発表 (2024.12.19)
・2024.12.06 欧州委員会 AI法における汎用AIモデル - Q&A (2024.11.20)
・2024.11.17 欧州委員会 汎用AI実践規範の最初のドラフトを発表 (2024.11.14)
・2024.10.30 欧州 AI法の調和標準の策定について...
・2024.09.08 欧州評議会 AI条約の署名開始... (2024.09.05)
・2024.08.05 欧州AI法が施行された... (2024.08.01)
・2024.07.19 ドイツ BfDI フランス CNIL オランダ AP EU AI法関連
・2024.07.16 EU 2024.07.12にAI法がEU官報に掲載された
・2024.05.22 EU 欧州理事会がAI法を承認...まもなく発効されますね...
・2024.03.19 欧州議会 AI法 (2024.03.13) とサイバーレジリエンス法を採択 (2024.03.12)
・2023.12.10 欧州委員会、欧州議会がAI法について政治的合意
・2022.12.08 EU理事会 AI法に関する見解を採択
・2022.09.30 欧州委員会 AI責任指令案
・2021.12.05 欧州理事会 AI法改正案を欧州議会に提出
・2021.08.08 EU議会 BRIEFING 人工知能法 at 2021.07.26
・2021.04.24 欧州委員会がAIへの規制を提案 → 欧州データ保護官は歓迎するけど、公共空間での遠隔生体認証についての規制も入れてね
« 経済産業省 情報セキュリティ監査基準と情報セキュリティ管理基準(令和7年改訂版)(2025.08) | Main | 米国 司法省 サイバー犯罪計画で使用される偽造身分証明書を販売するオンラインマーケットプレイスを押収 (2025.08.28) »

Comments