G7 - AI Development Guidelines

Historical Context and the Hiroshima Summit

The G7 Hiroshima AI Process (HAIP) was formally established by the Group of Seven (G7) leaders during the Hiroshima Summit in May 2023. This initiative was launched as a direct response to the unprecedented and rapid advancement of generative artificial intelligence and foundation models, which occurred throughout late 2022 and early 2023. Recognizing that these technologies offer transformative potential for economic growth and scientific discovery alongside significant risks to democratic institutions, human rights, and global security, the G7 leaders mandated their Digital and Tech Ministers to establish a forum for international discussion. The process aimed to foster a common understanding of the opportunities and challenges posed by advanced AI systems and to develop a set of international rules that could serve as a foundation for inclusive global governance. The HAIP represents a landmark effort by major economies—Canada, France, Germany, Italy, Japan, the United Kingdom, the United States, and the European Union—to align their regulatory and policy approaches, ensuring that innovation does not come at the expense of safety or shared democratic values. The Japanese G7 Presidency played a crucial role in facilitating this dialogue, emphasizing the need for 'Trustworthy AI' that is human-centric and respects the rule of law.

Core Definitions and Scope of Advanced AI

The Hiroshima AI Process introduces and relies upon several critical definitions to delineate its scope and application. Central to the framework is the term 'Advanced AI Systems,' which refers to the most sophisticated artificial intelligence technologies currently available, specifically highlighting foundation models and generative AI. These systems are characterized by their broad utility and ability to perform a wide range of tasks, which distinguishes them from narrow AI applications. By focusing on 'advanced' systems, the G7 targets the frontier of AI development where the potential for systemic risk is highest. The framework also defines 'Organizations' broadly to include any entity—whether in the private sector, academia, civil society, or the public sector—that is involved in the design, development, deployment, or provision of these advanced systems. This inclusive definition ensures that the responsibilities outlined in the principles apply across the entire AI lifecycle and ecosystem. Furthermore, the documents emphasize the concept of 'Trustworthy AI,' which is defined through adherence to the rule of law, human rights, due process, diversity, fairness, and non-discrimination. The framework explicitly states that AI should be 'human-centric,' meaning it must be designed and used in a way that respects human dignity and autonomy. Another key term is 'AI Actors,' which encompasses all individuals and entities involved in the AI lifecycle, from data curators and model developers to end-users and deployers.

The 11 Guiding Principles for Organizations

The core of the Hiroshima AI Process consists of 11 Guiding Principles that apply to organizations developing advanced AI systems. These principles are designed to be high-level yet actionable, providing a roadmap for responsible innovation. Principle 1 requires organizations to take appropriate measures throughout the AI lifecycle to identify, evaluate, and mitigate risks, including performing rigorous 'red-teaming' and internal/external testing prior to deployment. Principle 2 focuses on identifying and mitigating patterns of misuse after the system has been placed on the market, requiring developers to maintain an active role in monitoring their products' societal impacts. Principle 3 mandates that organizations publicly report their systems' capabilities, limitations, and domains of appropriate and inappropriate use, fostering transparency. Principle 4 encourages responsible information sharing among organizations and with governments regarding safety incidents and emerging risks. Principle 5 emphasizes the development and implementation of robust governance and risk management policies. Principle 6 calls for the investment in and implementation of robust security controls, including physical security and cybersecurity. Principle 7 is particularly notable, as it calls for the development and deployment of reliable content authentication and provenance mechanisms, such as watermarking, to help users identify AI-generated content. Principle 8 focuses on prioritizing research to mitigate societal, safety, and security risks. Principle 9 encourages the development of international technical standards. Principle 10 focuses on protecting personal data and intellectual property, while Principle 11 addresses the use of AI to tackle global challenges like the climate crisis and health emergencies.

The International Code of Conduct for Developers

The International Code of Conduct for Organizations Developing Advanced AI Systems serves as the operational arm of the Guiding Principles. While the principles provide the 'what,' the Code of Conduct provides the 'how' for organizations that choose to voluntarily commit to these standards. The Code is intended to be a 'living document,' capable of evolving alongside the technology it seeks to govern. It provides a set of voluntary commitments that organizations can publicly pledge to follow, creating a market signal of trustworthiness and responsibility. The Code emphasizes that while the benefits of advanced AI are global, the responsibility for safety starts with the developers. It specifically calls for developers to be transparent about their safety testing methodologies and to share the results of these tests with relevant stakeholders. By creating a standardized set of expectations, the Code of Conduct aims to prevent a 'race to the bottom' where companies sacrifice safety for speed. Instead, it encourages a 'race to the top' where adherence to G7 standards becomes a competitive advantage. The Code also highlights the importance of multi-stakeholder collaboration, urging developers to work with civil society and academia to understand the broader impacts of their models on marginalized communities and democratic processes.

Risk Management and Red-Teaming Protocols

A cornerstone of the HAIP is the emphasis on proactive risk management and the use of 'red-teaming' as a standard practice for advanced AI developers. Red-teaming involves the systematic and adversarial testing of AI models to identify vulnerabilities, biases, and potential for misuse before the models are released to the public. The HAIP specifically identifies several high-priority risk areas that must be addressed through these protocols. These include the potential for AI to be used in the development of chemical, biological, radiological, and nuclear (CBRN) weapons, as well as its potential to enhance cyberattacks or facilitate large-scale disinformation campaigns. Organizations are expected to conduct these tests not just once, but iteratively throughout the development process. The results of these tests should inform the deployment decisions, and if a model presents unmanageable risks, the principles suggest that it should not be released. This focus on 'safety by design' is intended to shift the burden of proof onto the developers, requiring them to demonstrate that their systems are safe for public use. The HAIP also encourages the use of external, independent auditors and researchers to verify the safety claims made by developers, adding a layer of objective oversight to the voluntary commitments.

Transparency, Disclosure, and Public Accountability

Transparency is a recurring theme throughout the Hiroshima AI Process, viewed as essential for building public trust and enabling effective oversight. Principle 3 and the corresponding sections of the Code of Conduct require organizations to provide detailed documentation about their models. This includes 'transparency reports' that describe the data used for training, the intended use cases of the model, and the known limitations or failure modes. Such disclosures are intended to help downstream deployers and end-users make informed decisions about how to use the technology safely. Furthermore, the HAIP calls for organizations to be transparent about their safety policies and the internal governance structures they have in place to manage AI risks. This includes identifying the individuals or committees responsible for safety decisions. Public accountability is also fostered through the requirement for organizations to report safety incidents and vulnerabilities to a centralized or coordinated body. By sharing information about what went wrong, the entire AI ecosystem can learn and improve, preventing the same mistakes from being repeated across different organizations. This collective learning approach is seen as vital for managing the systemic risks associated with foundation models that are integrated into thousands of different applications.

Content Authentication, Watermarking, and Disinformation

In the context of generative AI, the HAIP places significant emphasis on the integrity of the information ecosystem. Principle 7 specifically addresses the risks of AI-generated content being used to spread disinformation, manipulate public opinion, or facilitate fraud. To combat these threats, the G7 calls for the development and adoption of reliable content authentication and provenance mechanisms. This includes technical solutions like digital watermarking, metadata tagging, and cryptographic signatures that allow users to verify whether a piece of content—be it text, image, audio, or video—was created or modified by an AI system. The HAIP recognizes that no single technical solution is a silver bullet, and therefore encourages a multi-layered approach to content integrity. This includes educating the public about the existence of AI-generated content and encouraging platforms to implement clear labeling. The framework also urges organizations to support research into detection tools that can identify synthetic media even when provenance information is missing. By prioritizing these measures, the G7 aims to protect the 'right to truth' and ensure that the digital environment remains a space for authentic human discourse, particularly during sensitive periods such as elections.

Data Protection, Privacy, and Intellectual Property Rights

The Hiroshima AI Process acknowledges that the development of advanced AI systems is heavily dependent on vast amounts of data, which raises significant concerns regarding privacy and intellectual property (IP). Principle 10 explicitly requires organizations to implement robust data governance policies that respect the privacy of individuals and the rights of data subjects. This includes complying with existing data protection laws, such as the GDPR in Europe, and ensuring that personal data is collected and processed fairly and transparently. Organizations are also encouraged to use privacy-enhancing technologies (PETs) to minimize the risks of data leakage or re-identification. Regarding intellectual property, the HAIP emphasizes the need for organizations to respect copyright and other IP rights during the training of their models. This includes being transparent about the datasets used for training and providing mechanisms for rights holders to exercise their rights. The framework recognizes the ongoing global debate regarding the intersection of AI and copyright law and encourages organizations to participate in these discussions constructively. By balancing the need for data with the protection of individual and corporate rights, the HAIP seeks to create a sustainable and legally sound foundation for the AI economy.

Global Interoperability and the Friends Group

International alignment is a primary goal of the Hiroshima AI Process, as the G7 recognizes that AI governance cannot be effective if it is limited to a small group of countries. To achieve this, the G7 has engaged in extensive outreach to non-G7 partners, particularly in the 'Global South' and among emerging economies. This is facilitated through the 'Hiroshima AI Process Friends Group,' which provides a platform for broader participation and feedback. The G7 also coordinates closely with the United Nations, supporting the work of the UN Secretary-General’s High-Level Advisory Body on AI and contributing to the development of the Global Digital Compact. The HAIP also emphasizes the importance of 'interoperability' between different governance systems. This means that while different countries may choose different legal paths—some opting for binding regulation and others for voluntary guidelines—their underlying goals and standards for AI safety should be compatible. The G7 promotes this alignment through the development of shared technical standards and common definitions. By working with international standards-setting organizations, the G7 helps to ensure that the 'rules of the road' for AI are technically consistent worldwide. This international alignment is seen as essential for protecting the global digital economy and ensuring that the benefits of AI are shared equitably while its risks are managed collectively.

Implementation Timeline

Adoption and Endorsement Status

Key Stakeholders and Knowledge Partners

Sources and References