Hiroshima AI Behavior Guidelines: Complete English translation of Eleven Actions

Versions and Reading Instructions

This page presents the full substantive English translation of the "Hiroshima Process: International Behavioral Guidelines for the Development of Advanced AI Systems," as of October 30, 2023, corresponding to pages 1–8 of the original document. The original text was developed by the G7 Hiroshima AI Process and issued by the European Commission. It provides voluntary guidance and does not constitute equivalent to any current mandatory legal requirements in various regions.

The following retains the scope of the original document, the preface, and the eleven actions. The legal obligations assumed by an organization shall be determined according to the applicable regime. The word "should" in the translation expresses the advisory and commitment tone of the original guidelines, without claiming to be an official official translation.

Preface: Applicable Objects, Lifecycle, and Risk Methods

This guidance is based on the "International Principles for Organizing Development of Advanced AI Systems" and aims to promote global safety, reliability, and trustworthiness in AI. It provides voluntary action guidelines for organizations developing the most advanced AI systems, including state-of-the-art foundation models and generative AI systems; herein, they are collectively referred to as "advanced AI systems."

Organizations should take these actions based on risk. Entities that can support this framework include, but are not limited to, academic institutions, civil society, the private sector, and public sector entities.

This action list does not exhaust all possible scenarios. It is a living, evolving document that responds to recent advances in advanced AI, building upon the OECD AI Principles, to help harness the benefits of technology and address associated risks and challenges. Organizations should apply these actions throughout the entire lifecycle, covering design, development, deployment, and use as appropriate and relevant.

The document will be reviewed and updated as needed, including through ongoing inclusive multi-stakeholder consultations, to ensure its relevance and responsiveness to rapidly evolving technologies. Different jurisdictions may adopt their own approaches to implement these actions in different ways.

During periods when governments develop more enduring or detailed governance and regulatory frameworks, the guidelines call for organizations to engage in consultations with relevant stakeholders and act in a risk-based manner. The guidelines also commit to consulting with the OECD, GPAI, and other stakeholders to propose monitoring tools and mechanisms that help organizations take responsibility for the implementation of actions. Organizations are also encouraged to support the development of effective monitoring mechanisms by sharing best practices.

Moreover, organizations should establish an internal AI governance structure and policies, including self-assessment mechanisms, to provide a foundation of responsibility and accountability for actions and AI development.

While leveraging innovative opportunities, the design, development, and deployment of advanced AI should respect the rule of law, human rights, due process, diversity, fairness and non-discrimination, democracy, and a human-centered approach.

Organizations should not develop or deploy advanced AI in ways that are unacceptable, such as undermining democratic values, causing severe harm to individuals or communities, fostering terrorism, promoting criminal misuse, or posing significant risks to safety, security, and human rights.

States must fulfill their international human rights law obligations by fully respecting and protecting human rights; private sector activities should comply with international frameworks such as the United Nations Guiding Principles on Business and Human Rights and the OECD Due Diligence Guidance for Responsible Business.

The following eleven actions should be carried out in a manner commensurate with the associated risks.

1. Identify, assess, and mitigate risks throughout the AI lifecycle

Take appropriate measures to identify, assess, and mitigate risks throughout the development process of advanced AI systems, as well as during deployment and before market release.

This includes the use of diverse internal testing and independent external testing, combined with different evaluation methods such as red team testing, and the implementation of appropriate mitigations for identified risks and vulnerabilities. Testing and mitigation efforts should strive to ensure that the system is trustworthy, secure, and verifiable throughout its lifecycle, avoiding unreasonable risks. To support testing, developers should ensure that datasets, processes, and decisions throughout the system development lifecycle are traceable. These measures should be documented and supported by regularly updated technical documentation.

Testing should take place in a secure environment and be conducted at multiple checkpoints throughout the lifecycle, especially before deployment and market release. Testing is used to identify risks and vulnerabilities, helping to determine how to address unforeseen or intentional security, safeguards, social, and other risks.

When designing and implementing tests, the organization commits to addressing the following risks to an appropriate extent:

Chemistry, biology, radiological and nuclear risks : Advanced systems may lower the threshold for actors, including non-state actors, to develop, design, acquire, or use weapons.
Offensive cyber capabilities : The system may aid in identifying, exploiting, or using vulnerabilities for action. At the same time, it should be noted that some of these capabilities could also have defensive applications and are suitable for retention within the system.
Health or safety hazards : Includes interactions between systems and impacts from tool usage, such as the ability to control physical systems or disrupt critical infrastructure.
Model replication risk : Risks associated with creating a self-copy of a model, "self-replication," or training other models.
Societal, individual, and community risks : The system or model may generate harmful biases and discrimination, or lead to violations of applicable laws, including privacy and data protection regulations.
Threats to democratic values and human rights : for example, spreading false information or infringing on privacy.
Chain reaction : A particular event might trigger significant negative cascading effects, impacting the entire city, the entire event sector, or the entire community.

Fosters collaboration with cross-industry stakeholders to assess and implement mitigation measures, particularly for systemic risks.

Organizations making these commitments should also work to advance research and investment in: the safety, safeguards, bias, and misinformation of advanced AI, fairness, explainability and interpretability, transparency, and the robustness and trustworthiness of systems in the face of misuse.

Two. Post-deployment identification and mitigation of vulnerabilities, incidents, and abuse patterns

After deployment and market introduction, organizations should use the system as expected, based on risk levels, and continuously monitor vulnerabilities, incidents, emerging risks, and misuse, taking appropriate actions accordingly.

For example, vulnerability rewards, competitions, or awards can be used to encourage third parties and users to responsibly identify and report issues after deployment. Organizations should also appropriately document reported incidents and collaborate with other stakeholders to mitigate identified risks and weaknesses.

In cases where a reporting mechanism is appropriate, it should be accessible to diverse stakeholders, not limited to feedback from a specific group.

III. Public Reporting Capabilities, Limitations, and Appropriate and Inappropriate Uses

Organizations should publicly disclose the capabilities, limitations, and appropriate and inappropriate use cases of advanced AI systems to ensure sufficient transparency and enhance accountability.

A transparent report containing substantive information should be issued with every major new version release. Reports, user guides, and related technical documentation should be kept up to date as needed, for example including:

What assessments were conducted regarding potential safety, security, social, and human rights risks?
What capabilities does the model or system have, and which key performance limitations affect its applicable domains?
Discussion and evaluation of safety and societal impacts and risks, such as harmful biases, discrimination, threats to privacy or personal data protection, and effects on fairness.
Red team test results evaluating whether a model or system can transition from the development phase to subsequent stages.

Reports should be clear and understandable, enabling appropriate deployers and users to interpret model or system outputs and use the system correctly. Transparency reports should also be supported by solid documentation processes, such as technical documentation and user guides, rather than merely promotional statements.

Four. Responsiblely sharing information and reporting incidents

Organizations developing advanced AI should collaborate with the industry, government, civil society, and academia to promote responsible information sharing and incident reporting.

Appropriate information that may be shared includes, but is not limited to, evaluation reports, safety and assurance risks, anticipated or unanticipated capabilities, and attempts to circumvent safeguards throughout the AI lifecycle.

Organizations should establish or participate in appropriate mechanisms to develop, improve, and adopt common standards, tools, mechanisms, and best practices as needed to ensure the safety, assurance, and trustworthiness of advanced AI.

Appropriate and relevant documentation and transparency are essential throughout the entire lifecycle, particularly for systems that could pose significant risks to safety and society.

The organization should also collaborate with other organizations throughout the lifecycle to share and report relevant information with the public to advance safe, resilient, and trustworthy development; it should also, as appropriate, work with relevant public authorities and share such information. The reporting process should protect intellectual property rights.

Five. Develop, implement, and disclose AI governance and risk management policies

Policies should be based on a risk-based approach, including privacy policies and risk mitigation measures.

The organization should establish appropriate internal mechanisms to develop, disclose, and implement risk management and governance policies. For example, through accountability and governance processes, risks should be identified, assessed, prevented, and addressed throughout the lifecycle where feasible.

Appropriate disclosures of a privacy policy should be made, explaining how personal data, user prompts, and advanced system outputs are processed. Organizations should establish and publicly disclose governance policies and implementation mechanisms based on a risk-based approach, and ensure accountability and governance processes are evaluated and mitigate risks throughout the lifecycle.

The risk management policy should apply the risk management framework at appropriate and relevant times throughout the entire lifecycle, addressing a variety of risks associated with systems, and be regularly updated.

The organization should also establish policies, procedures, and training to ensure staff understand their responsibilities and the organization's risk management practices.

Six、Invest and implement sound security controls

Physical security, network security, and insider threat protection are required throughout the entire lifecycle.

Measures may include protecting model weights, algorithms, servers, and datasets, for example, implementing operational security measures for information security, as well as appropriate network and physical access controls.

Organizations should assess cybersecurity risks, establish cybersecurity policies, and implement adequate technical and organizational solutions to ensure the cybersecurity of advanced AI aligns with specific circumstances and relevant risks.

Model weights should be stored and used in appropriately secured and access-restricted environments to reduce risks of unauthorized disclosure and unauthorized access. This also includes establishing vulnerability management processes, regularly reviewing security measures to maintain high standards and ensure they remain effective against current threats.

Organizations should also establish robust internal threat detection programs whose level of protection is commensurate with the value of the most valuable intellectual property and commercial secrets, for example, by restricting access to proprietary and non-public model weights.

Seven、Establish a reliable content verification and source mechanism

When technically feasible, develop and deploy reliable mechanisms such as watermarks that enable users to identify content generated by AI.

For content generated by advanced AI systems within an organization, a content authentication and provenance mechanism should be provided whenever feasible and technically viable. Provenance information should include the identifier of the service or model that generated the content, but need not include user information.

Organizations should also strive to develop tools or APIs that allow users to determine whether content was generated by their advanced AI systems, for example, through watermarks. Organizations should collaborate and invest in research as needed to advance this field.

Moreover, other mechanisms such as labels and disclosures should be considered, and where feasible and appropriate, people should be made aware that they are interacting with an AI system.

Eight. Prioritize research and investment in effective risk mitigation

Prioritize research aimed at alleviating social, safety, and assurance risks, and allocate investment in effective mitigation measures.

This includes conducting, jointly carrying out, and investing in research on safety, assurance, and trust, addressing critical risks, and developing appropriate mitigation tools.

Prioritizes research and maintenance of democratic values, respect for human rights, protection of children and vulnerable groups, safeguarding intellectual property and privacy, and avoiding harmful biases, misinformation, deliberate false information, and information manipulation. Also commits to investing in appropriate tools and proactively managing risks associated with advanced systems, including environmental and climate impacts, to realize the benefits of technology.

Organizations are encouraged to share risk mitigation research and best practices.

9. Prioritize addressing global major challenges

Prioritize the development of advanced AI capable of addressing global grand challenges, particularly including but not limited to the climate crisis, global health, and education.

These efforts should support the United Nations Sustainable Development Goals, promoting AI development aligned with global common interests.

Organizations should responsibly maintain trustworthy and human-centered AI, supporting digital literacy education and training for the public, including students and workers, enabling them to benefit from advanced AI usage and helping individuals and communities better understand the nature, capabilities, limitations, and impacts of technology.

Organizations should collaborate with civil society and community groups to identify priority challenges, develop innovative solutions, and address the world's most pressing issues.

Ten. Promote international technical standards

Promote the development of international technical standards and adopt them when appropriate.

Organizations are encouraged to participate in international technical standards and best practices, including watermarking standards; collaborate with standard-setting organizations when developing test methods, content authentication and provenance mechanisms, cybersecurity policies, and public reporting measures.

Particularly, efforts should be made to develop international technical standards and frameworks with interoperability to help users distinguish between AI-generated and non-AI-generated content.

Eleven. Implement appropriate data input and rights protection measures

The organization should appropriately manage data quality, including training data and data collection, to mitigate harmful biases.

Measures may include transparency, privacy-protecting training techniques, and testing or fine-tuning to ensure the system does not leak confidential or sensitive information.

Organizations should implement appropriate safeguards to respect privacy and intellectual property rights, including those related to copyright-protected content, and comply with applicable legal frameworks.

Reference and Extended Reading

Original work: *Hiroshima Process International Code of Conduct for Organizations Developing Advanced AI Systems*, October 30, 2023. Source and rights note: European Union, issued by the European Commission's Digital Strategy. This page includes a English translation, chapter division, and terminology standardization; does not contain third-party images from external photo libraries, and does not imply official recognition by the original institution.

EU Commission webpage and original document : English, verify the original text and version against eight pages.
EU Digital Data Reuse Guidelines : English, indicate source attribution, modifications, and boundaries regarding third-party materials.
CC BY 4.0 : License notice. This translation is provided under this license; conditions for other content on the website are determined separately.

Hiroshima AI Behavior Guidelines: Full English translation of Eleven Actions

Clarify any unclear points.