2.1 A short history of AI governance

Following the deep-learning breakthroughs of the early 2010s, AI adoption surged across a myriad of industries and sectors (Brynjolfsson & McAfee, Reference Brynjolfsson and McAfee2014; LeCun et al., Reference LeCun, Bengio and Hinton2015; Bharadiya et al., Reference Bharadiya, Thomas and Ahmed2023). This rapid integration process brought to light a multitude of potential risks associated with deploying AI. From fatal accidents involving autonomous vehiclesFootnote ⁹ to discriminatory hiring practices by AI algorithms (Andrews & Bucher, Reference Andrews and Bucher2022), the real-world consequences of AI development have become increasingly evident. Furthermore, the manipulation of financial markets by algorithmic trading and the spread of misinformation on social media platforms (Ferrara, Reference Ferrara2024) highlight the broader societal concerns surrounding the technology’s integration across sectors.

Fueled by growing awareness of AI risks in the mid-and-late 2010s, national governments (including G7 members), international organizations, tech companies and nonprofits launched a wave of policy and principle publications. Prominent examples include the “Ethics Guidelines for Trustworthy AI” by the European Union (EU) in 2019,Footnote ¹⁰ the “Recommendation of the Council on Artificial Intelligence” by the OECD in 2019 (updated in 2024),Footnote ¹¹ and the “Recommendation on the Ethics of Artificial Intelligence” by the United Nations Educational, Scientific and Cultural Organization (UNESCO) in 2021.Footnote ¹² These publications emphasized pairing AI development with core values such as human rights, democracy and sustainability as well as key principles including fairness, privacy, safety, security, transparency and accountability.

While fundamental values and AI principles provide a crucial foundation to AI governance, translating them into implementable standards for AI systems remains a challenge, and addressing this challenge requires concrete and material guidance. Various initiatives have been undertaken at different levels to bridge this gap. At the national level, examples include the “AI Risk Management Framework”Footnote ¹³ (RMF) published by the National Institute of Standards and Technology (NIST) in the United States in January 2023, and Japan’s “AI Guidelines for Business” published by several ministries in April 2024.Footnote ¹⁴ On a supranational scale, leading examples include the 2023 AI Safety Summit’s “Emerging Processes for Frontier AI Safety”Footnote ¹⁵ and the G7’s HCOC – the latter being the focus of this article. Additionally, nongovernmental organizations such as the International Organization for Standardization (ISO) have contributed by issuing international standards on AI governance. The “AI Management System Standard ISO/IEC 42001”Footnote ¹⁶ was published in December 2023, specifying AI management system requirements. Another notable contribution to the risk management and stakeholder engagement field is the “Human Rights, Democracy, and the Rule of Law Assurance Framework for AI Systems” (HUDERAF), proposed by the Alan Turing Institute to the Council of Europe’s Ad Hoc Committee on Artificial Intelligence (Leslie et al., Reference Leslie, Burr, Aitken, Cowls, Katell and Briggs2022). Collectively, these diverse approaches underscore the ongoing efforts to transform abstract AI principles into a practical and implementable reality.

Despite this common direction, many published guidelines and principles for responsible AI development lack legally binding force, making them examples of “soft law.” While compliance with these documents helps companies with risk prevention strategies and forward-looking accountability measures, there are no guarantees or enforceability measures to ensure adherence to these standards. Thus, to advance stronger commitment to AI governance – in particular, addressing AI systems that pose high risks – there has been active movement to introduce regulations with legally binding force. For instance, the European Commission introduced the draft AI Act in 2021 (subsequently published in the Official Journal of the European Union on July 12, 2024),Footnote ¹⁷ focusing most of its compliance requirements on high-risk systems and even banning certain systems when the risks they present are deemed unacceptable.Footnote ¹⁸ Similarly, in 2022, Canada presented a legislative proposal, the Artificial Intelligence and Data ActFootnote ¹⁹ (AIDA), which focuses on establishing compliance requirements for high-impact AI applications. The United States has also seen a surge in legislative activity targeted at AI. As of August, 2024, over 105 draft bills have been introduced addressing AI,Footnote ²⁰ over 35 of which specifically target risk mitigation in AI applications.

The 2023 boom in foundation models presents a new layer of complexity to the already challenging landscape of AI governance. While conventional AI has faced issues such as limited explainability, diverse stakeholders and rapid evolution, foundation models expand the scope and reach of these challenges (Bommasani et al., Reference Bommasani, Liang, Hudson, Adeli, Altman, Arora, Von Arx, Bernstein, Bohg, Bosselut, Brunskill, Brynjolfsson, Buch, Card, Castellon, Chatterji, Chen, Creel, Davis, Demszky, Donahue, Doumbouya, Durmus, Ermon, Etchemendy, Ethayarajh, Fei-Fei, Finn, Gale, Gillespie, Goel, Goodman, Grossman, Guha, Hashimoto, Henderson, Hewitt, Ho, Hong, Hsu, Huang, Icard, Jain, Jurafsky, Kalluri, Karamcheti, Keeling, Khani, Khattab, Koh, Krass, Krishna, Kuditipudi, Kumar, Ladhak, Lee, Lee, Leskovec, Levent, Li, Li, Ma, Malik, Manning, Mirchandani, Mitchell, Munyikwa, Nair, Narayan, Narayanan, Newman, Nie, Niebles, Nilforoshan, Nyarko, Ogut, Orr, Papadimitriou, Park, Piech, Portelance, Potts, Raghunathan, Reich, Ren, Rong, Roohani, Ruiz, Ryan, Ré, Sadigh, Sagawa, Santhanam, Shih, Srinivasan, Tamkin, Taori, Thomas, Tramèr, Wang, Wang, Wu, Wu, Wu, Xie, Yasunaga, You, Zaharia, Zhang, Zhang, Zhang, Zhang, Zheng and Zhou2021).Footnote ²¹ The application of these systems in countless contexts and their ease of operation create an even more intricate risk environment. As a result, there has been a surge in global efforts to establish rules and foster international cooperation around foundation models. The EU AI Act,Footnote ²² for example, has incorporated provisions specifically related to “general-purpose AI” systems.Footnote ²³ Japan’s Liberal Democratic Party proposed the concept note for the Basic Law for the Promotion of Responsible AI in February 2024,Footnote ²⁴ which targets advanced foundational AI models with significant societal impact. Similarly, the Chinese government implemented the Interim Measures for the Administration of Generative Artificial Intelligence ServicesFootnote ²⁵ in August 2023, establishing specific requirements for models with “public opinion properties or the capacity for social mobilization.”Footnote ²⁶ Figure 1 shows the overall structure of AI governance and key documents related to each layer of governance.

Figure 1.

Overall structure of AI governance and key documents related to each layer.

The brief history of AI governance is characterized by a complex and multidimensional balancing act between innovation and regulation, rapidly advancing technology, and the integration of multivector interests – encompassing the technology industry, the general public and regulators. While these groups may have differing priorities, there is also growing recognition of the need for collaboration. Responses to AI risks have evolved: Nations and international bodies initially relied on soft-law principles and public–private collaborative efforts, whereas the current momentum is toward binding legislative action, with specific measures addressing advanced AI. Another crucial distinction is the regulatory scope, which can be generally categorized as comprehensive or sectorial. While the EU’s AI Act and Canada’s proposed AIDA encompass regulations that span across industries, Japan, the United Kingdom (UK) and the United States have indicated a policy direction that considers industry-specific AI regulations or focuses on powerful foundational models. Nonetheless, the regulatory emphasis in all of these instances is primarily on high-risk AI, aiming to strike an appropriate stability between fostering technological development and ensuring safety. G7 democracies, in particular, find common ground in core principles such as human rights and democratic values, grounding them in transparency, explainability and bias prevention and forming a common foundation for responsible AI development.Footnote ²⁷

2.2 Advancing international collaboration

This following subsection first (1) provides an overview of key international AI governance initiatives, including significant documents and declarations such as the G7’s HAIP Comprehensive Framework, the Bletchley Declaration, the UN’s “Governing AI for Humanity” report, and the Council of Europe’s AI Treaty. These documents highlight various efforts to establish global standards and frameworks for AI governance. Subsequently, (2) the discussion will examine the pivotal role of the G7’s framework in shaping global AI governance. The G7’s role in global AI policies is underscored by its active participation in major international initiatives and its significant economic, regulatory, and technological impact.

2.2.1 Key international AI initiatives

As countries make progress with AI rulemaking within their borders, international cooperation is also advancing.Footnote ²⁸ The G7 is one of the most impactful forums for such international coordination. During the May 2023 summit, G7 leaders committed to establishing the HAIP by the end of the year to foster collaborative policy development on generative AI.Footnote ²⁹ Within 6 months, the G7 digital and tech ministers had delivered the Comprehensive Framework.Footnote ³⁰ This framework prioritizes proactive risk management and governance, transparency and accountability across the AI life cycle.Footnote ³¹ Additionally, it emphasizes anchoring AI development in human rights and democratic values while fostering the use of advanced AI for tackling global challenges such as climate change, health care, and education.Footnote ³²

In November 2023, the AI Safety Summit held in the UK produced the “Bletchley Declaration,” a significant milestone in international AI collaboration.Footnote ³³ The declaration addresses crucial aspects of AI governance, such as the protection of human rights, transparency, explainability, fairness, accountability, human oversight, bias mitigation, and privacy and data protection.Footnote ³⁴ Additionally, it highlights the risks associated with manipulating or generating deceptive content.Footnote ³⁵ Endorsed by 29 countries and regions, the signatories encompass not only G7 and OECD nations but also partners from the Middle East, Africa, South America, Asia, and, notably, China.Footnote ³⁶ A second AI Safety Summit was held in Seoul in May 2024,Footnote ³⁷ which reiterated the anchoring point of safety, and highlighted inclusion and innovation as critical priorities for global convergence.Footnote ³⁸

The United Nations is also active in forming international AI governance principles. In December 2023, the UN AI Advisory Body issued the interim report “Governing AI for Humanity.”Footnote ³⁹ The report outlines a set of guiding principles and institutional roles designed to create a global AI governance framework, proposing essential considerations and actions to ensure that AI development and deployment serve the broader interests of humanity.Footnote ⁴⁰ These include principles such as inclusivity,Footnote ⁴¹ public interestFootnote ⁴² and the importance of aligning AI governance with data governance and promoting a data commons.Footnote ⁴³ Institutional functions highlighted in the report include assessing the future directions and implications of AIFootnote ⁴⁴; developing and harmonizing standards,Footnote ⁴⁵ safety and RMFsFootnote ⁴⁶; and facilitating the development, deployment, and use of AI for economic and societal benefit through international multi-stakeholder cooperation.Footnote ⁴⁷

In March 2024, the Council of Europe’s Ad Hoc Committee on Artificial Intelligence introduced the Framework Convention on Artificial Intelligence, Human Rights, Democracy and the Rule of Law (AI Treaty), a groundbreaking treaty on AI governance that sets a high bar on responsible AI development.Footnote ⁴⁸ The AI Treaty, adopted in May 2024,Footnote ⁴⁹ emphasizes the obligation of signatory nations (parties to the convention) to proactively ensure AI activities are aligned with human rights, democratic integrity and the rule of law. The treaty calls for comprehensive safeguards throughout the AI life cycle – including mechanisms for accountabilityFootnote ⁵⁰ and transparencyFootnote ⁵¹ – and introduces comprehensive RMFs.Footnote ⁵² Furthermore, it calls for robust remedies and procedural protective measures against rights violations,Footnote ⁵³ promotes rigorous risk and impact assessments,Footnote ⁵⁴ and delineates duties for international cooperation and implementation, focusing on nondiscrimination and rights protection.Footnote ⁵⁵

Nations participating in these initiatives vary. Figure 2 maps the structural involvement of various jurisdictions in the abovementioned international processes.

Figure 2.

The global AI governance landscape.

Note: The European Union (EU) is considered a non-enumerated member of the G7. Both the European Council and European Commission presidents participate in the leaders’ summit. The countries shown in blue indicate those participating in the Hiroshima AI Process Friends Group (as of December 31, 2024). The nations with dots (●), plus the G7 and EU members, are members or observers of the Council of Europe, the host organization of the AI Treaty.

3.1 Structure of the Comprehensive Framework

In response to the rapid development and global spread of advanced AI, the G7 nations launched the HAIP in May 2023 under Japan’s presidency.Footnote ⁶⁰ This international forum aims to establish common ground for responsible AI development and use. It focuses on fostering safe, secure and trustworthy AI by addressing key ethical issues, promoting collaboration on research and development, and encouraging international standards for a future where humanity benefits from AI advancements. Although the HAIP focuses on governance of advanced AI systems, the Comprehensive Framework avoids a rigid definition of this technology by providing the tentative definition of “the most advanced AI systems, including the most advanced foundation models and generative AI systems.”Footnote ⁶¹ This flexibility likely reflects a desire to adapt to future advancements in AI performance, functionalities, and deployment landscapes.

The Comprehensive Framework consists of four elements (see Figure 3). First, the OECD’s “G7 Hiroshima Process on Generative Artificial Intelligence”Footnote ⁶² serves as a background analysis of the opportunities and risks of advanced AI systems. Second, the “Hiroshima Process International Guiding Principles for All AI Actors”Footnote ⁶³ (HIGP) provides 12 general principles for designing, developing, deploying, providing and using advanced AI systems without providing detailed guidance. Third, the HCOCFootnote ⁶⁴ consists of a set of detailed instructions for the developers of advanced AI systems under the general principles the HIGP provides. Finally, the “project-based cooperation” on AI includes international collaborations in areas such as content authentication and the labeling of AI-generated content.

Figure 3.

Four elements of the HAIP Comprehensive Framework.

The following section summarizes the contents of the HIGP and HCOC.

3.2 Contents of the HIGP

The HIGP is a comprehensive set of values and best practices promoting responsible development and use of advanced AI on a global scale. It consists of 12 core principles that serve as a foundation for responsible AI governance. These principles closely mirror the values and approaches that G7 nations are already exploring for their individual AI governance frameworks.Footnote ⁶⁵ The analysis here suggests that the 12 principles may be divided into the following three groups (see Table 1):

1. 1. Risk management and governance: This group recommended actions to assess and mitigate risks associated with AI systems, ensuring they are reduced to a level that relevant stakeholders deem acceptable.

2. 2. Stakeholder engagement: This group recommended actions to ensure clear communication with and accountability to all relevant stakeholders.

3. 3. Ethical and societal considerations: This group recommended actions to ensure the development, deployment and usage of AI are in alignment with ethical standards and societal values.

Table 1.

Elements of the Hiroshima Process International Guiding Principles

Note: The numerals listed for each item correspond to the articles assigned in the HIGP and HCOC. The authors devised the abbreviations for the principles and their categorization.

3.3 Overview of the Code of Conduct

Building on 11 of the HIGP’s 12 core principles (excluding the trustworthy and responsible use of advanced AI),Footnote ⁶⁶ the HCOC translates these principles and materializes them into a more specific code of practice for organizations developing and deploying advanced AI systems. The HCOC provides a comprehensive road map for AI processes and risk mitigation, outlining general actionable items on the matters of risk management and governance, stakeholder engagement, and ethical considerations.Footnote ⁶⁷

3.3.3 Ethical and societal considerations

The HCOC establishes in items 8–10 a series of parameters to ensure AI is developed and deployed within the boundaries of human rights and democracy to address global challenges:

• • Research prioritization for societal safety: emphasizing collaborative research to advance AI safety, security and trustworthiness, focusing on key risks such as upholding democratic values, respecting human rights and protecting vulnerable groups

• • AI for global challenges: prioritizing development of advanced AI systems to address global challenges such as climate change, health and education, aligning with the UN Sustainable Development Goals

• • International technical standards: encouraging contribution to the development and use of international technical standards, including practices to promote transparency by allowing users to identify AI-generated content (e.g., watermarking), testing methodologies and cybersecurity policies

A detailed summary of the HCOC is presented in Table 2.

Table 2.

Summary of the Hiroshima Process International Code of Conduct

Note: The numerals listed for each item correspond to those assigned in the HIGP and HCOC. The authors devised the abbreviations for the principles and their categorization.

4.1 Status of AI governance in the G7 and HCOC as common guidance

The HCOC serves as a central reference point in the evolving global landscape of AI governance. This section provides insight into how HCOC aligns with the existing frameworks in G7 jurisdictions, including Canada, the EU, Japan, the UK and the United States. Next, the section contains an AI-focused overview of each jurisdiction’s regulatory status, identifies the documents that closely align with the HCOC’s structure and functionality, and evaluates their compatibility with the HCOC’s content. The summary of the analysis is shown in the “Annex.”

1. 1. Canada: Canada is in the process of formulating a comprehensive regulatory framework for AI under Bill C-27, known as AIDA.Footnote ⁷³ This legislation prioritizes risk mitigation for “high-impact” AI systems.Footnote ⁷⁴ Additionally, Canada has published a Voluntary Code of Conduct for Responsible Development and Management of Advanced Generative AI Systems,Footnote ⁷⁵ offering nonbinding guidelines for AI industry stakeholders.

2. 2. European Union: The EU has positioned itself at the forefront of AI regulation with the AI Act, published in July 2024.Footnote ⁷⁶ This legislation sets a robust and comprehensive framework for trustworthy AI development and implementation, emphasizing a risk-based regulatory approach.Footnote ⁷⁷ The AI Act mandates the development of codes of practice to guide its implementation, ensuring alignment with international standards as well as evolving technology and market trends.Footnote ⁷⁸

3. 3. Japan: Japan’s approach to AI governance emphasizes maximizing the positive societal impacts of AI and capitalizing on a risk-based and agile governance model.Footnote ⁷⁹ Taking a sector-specific approach, Japan seeks to promote AI implementation through regulatory reforms tailored to specific industries and markets, such as transportation, finance, and medical devices.Footnote ⁸⁰ This strategy includes updating more than 10,000 regulations or ordinances that require “analog” compliance methods, including requirements for paper documents, on-site periodic inspections and dedicated in-person staffing.Footnote ⁸¹ In addition, Japan launched the AI Guidelines for BusinessFootnote ⁸² as a voluntary AI risk management tool. The principles for advanced AI systems established in the HIGP are directly integrated into these guidelines, following Japan’s presidency of the G7 during the HAIP Comprehensive Framework drafting process.

4. 4. United Kingdom: The UK is developing a decentralized regulatory approach focusing on sector-specific guidelines, a pro-innovation stance and public–private collaboration through specialized AI institutions.Footnote ⁸³ While the UK is not currently enforcing a comprehensive AI law or drafting a central code of conduct, it emphasizes traditional AI governance principles such as safety, security, transparency, and fairness to inform its sector-driven regulations.Footnote ⁸⁴ The UK Department for Science, Innovation and Technology also published a practical guidance code in the form of the Emerging Processes for Frontier AI SafetyFootnote ⁸⁵ ahead of the UK AI Safety Summit. The summit culminated in the Bletchley Declaration, a shared commitment to safe and responsible AI development signed by 28 nations and the EU.Footnote ⁸⁶

5. 5. United States: The United States has adopted a decentralized, multitiered regulatory strategy for AI governance, with agencies overseeing sector-specific regulations.Footnote ⁸⁷ Key initiatives include the “Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence,”Footnote ⁸⁸ which directs sector-specific agencies to formulate regulations; the Risk Management Frameworks,Footnote ⁸⁹ developed by NIST to provide guidelines for risk assessment and management; the “White House’s Blueprint for an AI Bill of Rights,”Footnote ⁹⁰ outlining foundational principles for AI development; and nonbinding voluntary commitments for ensuring safe, secure and trustworthy AIFootnote ⁹¹ endorsed by companies such as Amazon, Anthropic, Google, Inflection, Meta, Microsoft, Nvidia and OpenAI, among others.

4.2 Achieving and enhancing regulatory interoperability: The HCOC as a reference point for AI governance development

The AI governance landscape across the G7 is complex and multifaceted. The EU has instituted robust and comprehensive regulations through its AI Act, and Canada is in the process of developing similar hard-law frameworks.Footnote ⁹² Conversely, the United States, Japan and the UK lean toward sector-specific and lighter-touch regulatory approaches.Footnote ⁹³ This diverse regulatory environment, marked by varying levels of stringency, scope and focus, poses challenges for global operations, requiring businesses to navigate a complex regulatory patchwork, as well as differing rights and obligations across G7 nations. The HCOC holds promise as a unifying mechanism, to bridge these regulatory disparities and promote interoperability.

The HCOC may be integrated into national regulations across G7 countries through various means, such as direct formal legal referencing or recognition, material content integration, and leveraging or harmonizing specific aspects of regulatory developments. Pathways for integration into the regulatory frameworks of the G7 jurisdictions include the following:

• • Canada: Overall, Canada’s Voluntary Code of Conduct specifically, and its regulatory trajectory generally, demonstrate alignment with the international conversation on ethical AI development and the HCOC’s principles.Footnote ⁹⁴ As AIDA evolves, it presents the potential to translate these principles into enforceable regulations, further solidifying Canada’s commitment to responsible AI advancement. Given that AIDA could likely address advanced AI systems specifically within its regulatory scope, this upcoming law opens a clear possibility to find common ground with the HCOC’s principles and functionality.

• • European Union: The EU AI Act mandates the development of codes of practice that complement its implementation.Footnote ⁹⁵ These codes of practice align with the HCOC’s focus, addressing practical aspects of responsible AI development. Furthermore, the EU acknowledges in the EU AI act that international standards should play a role in shaping these codes of practice,Footnote ⁹⁶ presenting an opportunity to materially integrate or formally reference the HCOC in the EU AI governance framework.

• • Japan: In February 2024, the Liberal Democratic Party proposed the concept note for the Basic Law for the Promotion of Responsible AI.Footnote ⁹⁷ The proposed legislation specifically targets advanced foundational AI models with significant societal impact. It requires model developers to adhere to seven key measures,Footnote ⁹⁸ including third-party vulnerability checks and the disclosure of model specifications. The requirements align with the voluntary commitments the White House has requested from U.S. companies.Footnote ⁹⁹ The HCOC could serve as a valuable reference point for implementation of these key measures, especially considering that the HCOC principles are already integrated into Japan’s AI Guidelines for Business.Footnote ¹⁰⁰

• • United Kingdom: Besides leading international discussions on AI governance through initiatives such as the Bletchley Declaration,Footnote ¹⁰¹ the UK is proactively formulating its own AI governance framework. According to “A Pro-innovation Approach to AI Regulation,”Footnote ¹⁰² the UK government is undergoing technical policy analysis on regulation and life-cycle accountability of capable general-purpose systems. It also commits to updating the Emerging Processes for Frontier AI, Safety,Footnote ¹⁰³ which is highly compatible with the HCOC, by the end of 2024. For these purposes, the UK is opting for collaborative private-public development through institutions such as the Digital Regulation Cooperation ForumFootnote ¹⁰⁴ and the AI Safety Institute.Footnote ¹⁰⁵ Considering the current institutional inertia and the stalled progress regarding its draft intellectual property code,Footnote ¹⁰⁶ the UK could leverage the HCOC and its international scope to inform these regulatory initiatives.

• • United States: The United States is in active development of its AI governance frameworks. The AI executive order has directed multiple agencies to deliver sector-specific guidance publications, and as of August 2024 there are more than 105 draft bills addressing AI, with over 35 focused on risk mitigation.Footnote ¹⁰⁷ Notably, after releasing RMF 1.0 in January 2023, NIST established the Generative AI Public Work GroupFootnote ¹⁰⁸ to spearhead development of a cross-sectoral AI RMF profile for managing the risks of generative AI models or systems.Footnote ¹⁰⁹ The HCOC’s emphasis on responsible risk management and governance aligns seamlessly with the United States’ principles-based trajectory and could fit into proposed risk mitigation legislation, positioning the HCOC as a crucial reference in shaping AI regulatory policy in the United States.

5.1 Terminology and definitions: Indexing a common vocabulary

The HCOC can serve as a foundation for a consistent definition or methodology for identification of terms for advanced AI systems governance, facilitating smoother regulatory implementation across jurisdictions. Future terminology consensus includes the following:

• • Bridge the terminology gap: The HCOC can endorse consistent definitions for streamlined regulatory implementation across jurisdictions, fostering a common understanding of critical concepts. This could be achieved by including a glossary of key terms with clear, agreed-upon definitions or by establishing methodologies for identifying and classifying AI systems based on the factors relevant to risk assessment. By establishing a common language, the HCOC can ease communication, regulatory certainty and business-sector collaboration across borders. Underscoring the importance of shared language around AI, the EU and the United States are currently in the development of 65 key terms “essential to understanding risk-based approaches to AI.”Footnote ¹¹⁰ Notably, even when common terminology has been developed (e.g., through the U.S.-EU Trade and Technology Council, OECD or the ISO), the definition of advanced AI systems is unclear, leaving the question of which criteria (e.g., floating point operations, quality and size of data set, or input and output modalities)Footnote ¹¹¹ should be used to determine advanced AI systems.

5.2 Risk management and governance: Building a common and robust framework

Effective risk management stands as a cornerstone of responsible development of advanced AI systems. The HCOC can significantly contribute to this endeavor by advocating for shared principles and best practices. Risk management cohesion across jurisdictions includes the following:

• • Identify and share security risks, particularly systemic risks: The HCOC can enhance its interoperability contribution by explicitly listing and addressing security risks, particularly those with systemic consequences. This can be achieved through a two-pronged approach. First, the HCOC can integrate a comprehensive list of typical AI risks common to advanced AI systems, such as AI hallucinations (generating inaccurate outputs), fake content generation (deepfakes), intellectual property infringement (copyrighted content integration in data sets), job market transformations due to automation, the environmental impact of AI systems, bias amplification based on training data and privacy concerns, among others. Case studies can be implemented through “project-based cooperation,” which constitutes the fourth element of the Comprehensive Framework. Second, the HCOC can establish a risk assessment framework to categorize AI systems based on their potential for harm. This framework could leverage existing models such as the EU AI Act’s categorization of general-purpose AI models with systemic risk and its classification rules for high-risk AI systems.Footnote ¹¹² By prioritizing systems with the greatest potential for systemic or high-impact issues, the HCOC can provide a clearer road map for identifying, understanding, and mitigating various risks.

• • Enhance clarity in the risk management process: The HCOC can encourage the development of standardized risk management policies tailored to specific AI applications. Future drafting can reference or draw insights from established RMFs, such as ISO/IEC 42001:2023 or NIST’s RMF – especially the RMF developed by the Generative AI Public Working GroupFootnote ¹¹³ in July 2024. Additionally, policies can incorporate learnings from other reputable sources to enhance clarity and comprehensiveness.

• • Develop standard data governance, risk management and information security policies: Establishing robust data protection protocols is essential for building trust and mitigating risks associated with AI development. The development of standardized policies can leverage established frameworks such as ISO/IEC 27001 and ISO/IEC 27002 or NIST’s Cybersecurity Framework, which provide a structured foundation adaptable to the unique risk landscape of the development of advanced AI systems.

• • Implement content authentication mechanisms: The HCOC can list reliable content authentication and provenance mechanisms to enable users to identify the originators of content or establish common labeling mechanisms to help users understand that AI has generated the content. These contributions could be based on input from the HAIP’s project-based cooperation. Authentication mechanisms can safeguard against misinformation and uphold democratic values and human rights by verifying data sources and outputs. However, it is imperative to balance these efforts with the protection of individual privacy, ensuring authentication processes do not compromise personal data. This balance is key to maintaining public trust and promoting the responsible and user-centric deployment of AI technologies.

5.3 Stakeholder engagement: Fostering transparency and accountability

Building trust in AI necessitates robust stakeholder engagement. A transparent and accountable AI development process fosters public confidence and encourages information sharing. Future pathways for stakeholder engagement include the following:

• • Establish standardized formats for transparency reports: The HCOC can promote the adoption of standardized formats for transparency reports. By consolidating best practices and identifying common risks, the HCOC can offer a template for companies to self-assess and disclose relevant information consistently across jurisdictions. A potential model for standardized formatting pursuant to transparency reports is the UK Algorithmic Transparency Recording Standard.Footnote ¹¹⁴ Standardization would enable companies to have uniform international disclosure criteria, enhancing cross-border cohesive reporting and auditing consistency as well as allowing the public to better understand the development and operation of AI systems.

• • Define clear formats for incident sharing: Encouraging adoption of clear incident-sharing formats can facilitate the exchange of information about security breaches, biases or unintended consequences observed in deployed AI systems. This collaborative approach to sharing and learning from incidents enables stakeholders to develop effective mitigation strategies, ultimately enhancing the safety and reliability of AI technologies.

5.4 Ethical and societal considerations: Upholding the rule of law, human rights and core democratic values

The G7, a group of leading democracies, has a unique opportunity to shape the global conversation around responsible AI development. The HCOC, as an initiative stemming from this group, can play a crucial role in ensuring AI development aligns with the ethical and societal considerations that underpin democratic values and secure human rights in AI development and implementation. Potential pathways to prioritizing these principles include the following:

• • Reinforce the primacy of rule of law, human rights and democratic principles: The HCOC already champions these core values and emphasizes human-centric design. However, there is room for further enhancement and substantiation for practical application. For instance, the HCOC could enhance its guidance on how organizations should foster research and AI development that prioritizes the protection of fairness, privacy and intellectual property rights while also tackling global challenges such as climate change, health and education. Rather than providing detailed descriptions itself, the HCOC could reference other international agreements or widely recognized standards. Furthermore, the HCOC could strengthen democratic principles and the rule of law by highlighting due safeguards for freedom of expression, ensuring AI does not minimize dissent or impose undue restrictions on information access, guaranteeing a right to remedy for individuals adversely affected by AI and promoting transparency and accountability in AI decision-making processes. Enhancing human-centricity could involve advocating for effective oversight in high-risk applications, providing individuals with explanations regarding AI-driven decisions affecting them, and promoting inclusive design that caters to the diverse needs and perspectives of various populations to ensure equitable AI benefits.

5.5 Further areas for exploration

The HCOC can play a key role in exploring several critical areas for further development in responsible AI:

• • Acknowledge special considerations for government use of AI: The HCOC can play a pivotal role in delineating special considerations for government use of AI, ensuring governmental powers in AI deployment are appropriately circumscribed and limited. Drawing inspiration from the AI TreatyFootnote ¹¹⁵ and leveraging principles from the OECD Declaration on Government Access to Personal Data Held by Private Sector Entities,Footnote ¹¹⁶ the HCOC can establish clear guidelines that emphasize due process in developing and deploying advanced AI systems by the public sector, such as legal basis, legitimate aims, oversight, and redress, in addition to shared principles such as privacy, transparency and accountability. By aligning with these principles, the HCOC can become a democratic referent, and governments can leverage the power of AI responsibly while mitigating potential risks and fostering public trust.

• ● Harmonize full life cycle regulatory approaches: The HCOC can explore the potential for incorporating best practices from various jurisdictions’ regulations. This could involve elements such as certification mechanisms, robust oversight mechanisms and iterative audit controls.

  • ∘ Certification mechanisms: The HCOC can establish a framework for certification and registration mechanisms for high-risk advanced AI systems. This system could ensure rigorous evaluation throughout the life cycle of high-risk and advanced AI systems, from pre-market integrative assessments to ongoing post-market analyses and compliance reviews. The HCOC could define risk categories and establish criteria for the need for certification.

  • ∘ Oversight methodologies: The HCOC can emphasize the importance of effective oversight in AI systems to mitigate potential harm and address incidents effectively. In some cases, human involvement in critical AI processes is necessary, while in other cases machines can detect risks much faster and more precisely than humans. The HCOC could propose guidelines about whether to prioritize human judgment and intervention, especially in high-risk AI applications, ensuring a balance between automation and human control.

  • ∘ Audit mechanisms: The HCOC can extend procedural cohesion beyond AI implementation by establishing common processes for iterative audits, ensuring continuous monitoring and evaluation of AI systems’ compliance with established principles and guidelines. By considering and potentially adapting existing frameworks, such as the UK Guidance on the AI Auditing Framework,Footnote ¹¹⁷ the HCOC can equip organizations with practical tools for ongoing evaluations. These iterative audits would allow for continuous improvement and ensure AI systems remain aligned with responsible development principles throughout their life cycle.

• • Establish means for redress: The HCOC could expand discussions about redress for harms caused by advanced AI systems. This could involve exploring access to remedies and explanations for individuals affected by AI decisions in areas as diverse as copyright and intellectual property to judicial processing. As AI plays a growing role in judicial decision-making, for example, developing specific appeal mechanisms for harms caused by AI-based judicial decision making may become also crucial. The HCOC could encourage developers and deployers of advanced AI systems to provide appropriate dispute resolution mechanisms to users and harmed parties. Furthermore, to make victim relief more effective, G7 members could discuss shifting the burden of proof of damages or causal links and establishing accessible, fast and low-cost dispute resolution mechanisms for damages caused by advanced AI systems.

• • Foster shared responsibility in the AI ecosystem: The HCOC addresses developers of advanced AI systems only.Footnote ¹¹⁸ However, its scope could expand in the future to other actors within the AI value chain, such as deployers and users of advanced AI systems. In addition, it is important to examine how to distribute responsibility and liability among stakeholders, ensuring all parties are accountable for their respective roles in potential harms.

By focusing on these key areas, the HCOC can evolve into a powerful tool for facilitating a more cohesive and effective approach to AI governance on a global scale. The HCOC’s dynamic nature positions it to bridge the gap between diverse national frameworks, fostering a future of responsible AI development for the G7 nations and beyond.