The Hiroshima AI Process — The G7's Guiding Principles and Code of Conduct for Generative AI

The Hiroshima AI Process is a framework for international rules on generative AI, launched in May 2023 following the G7 summit in Hiroshima under Japan's presidency. On 30 October 2023, G7 leaders issued a statement welcoming the Hiroshima Process International Guiding Principles and the International Code of Conduct for Organizations Developing Advanced AI Systems, and called on developers to commit to applying the code. In December these were folded into the Hiroshima AI Process Comprehensive Policy Framework, which G7 leaders endorsed. Japan's Ministry of Internal Affairs and Communications (MIC) describes it as the first international policy framework made up of guiding principles and a code of conduct. None of it is legally binding; it relies on voluntary action by companies.

The May communiqué: discuss it by year-end

The G7 Hiroshima Leaders' Communiqué of 20 May 2023 said the opportunities and challenges of generative AI needed to be taken stock of immediately, and tasked ministers with establishing the "Hiroshima AI process" through a G7 working group, in cooperation with the OECD and the Global Partnership on AI (GPAI), for discussions on generative AI by the end of the year. Suggested topics included governance, protection of intellectual property including copyright, promotion of transparency, responses to foreign information manipulation including disinformation, and responsible use of the technology.

This was less than six months after the launch of ChatGPT and about two months after GPT-4.

Six months of work

30 October was also the day the US issued the Biden executive order. The G7 statement closed by looking forward to the UK's AI Safety Summit two days later.

Four elements of the framework

According to the ministers' statement of 1 December, the Comprehensive Policy Framework has four elements:

1. The OECD's report towards a G7 common understanding on generative AI
2. The Hiroshima Process International Guiding Principles for All AI Actors
3. The Hiroshima Process International Code of Conduct for Organizations Developing Advanced AI Systems
4. Project-based cooperation on AI

The developer-facing principles and code came first, in October; December added principles addressed to all AI actors. According to an analysis by CSIS, the all-actor principles comprise 12 principles, and the code of conduct turns 11 of them (all except the responsible use of advanced AI) into concrete actions for developers.

The 11 actions of the code of conduct

Action 1 lists risks that testing should pay particular attention to, including lowering barriers to chemical, biological, radiological and nuclear (CBRN) weapons, offensive cyber capabilities, self-replication, harmful bias and discrimination, and threats to democratic values through disinformation. The code also states that developing or deploying advanced AI in ways that undermine democratic values, facilitate terrorism or pose substantial risks to human rights is not acceptable.

The Friends Group and the reporting framework

Two mechanisms extended the process beyond the G7. The Friends Group is a voluntary grouping of countries that support the spirit of the process; 49 countries and regions backed it at launch. MIC's list shows 69 countries plus the EU as of August 2026, including emerging economies in Africa and Asia. A "Partners Community" of companies and international organisations supports it, including Anthropic, Amazon, Fujitsu and NEC.

The reporting framework asks AI developers to report publicly on how they apply the code. It began operating in February 2025; responses from 19 organisations were published in April, and 31 organisations responded to the second version published in October 2026. MIC cites participants' reports of benefits such as better coordination between internal teams and a benchmark for comparing their governance with international standards.

Assessment: the strengths and limits of soft law

Nothing in the Hiroshima documents carries a penalty. A May 2024 analysis for CSIS by Hiroki Habuka and David U. Socol de la Osa observed that many responsible-AI principles lack legally binding force—they are "soft law"—with no guarantees or enforcement to ensure adherence. It nevertheless argued that the G7 could anchor interoperability between national rules: its members share democratic values, accounted for about 26.4% of global GDP in 2023, and are home to most companies developing advanced AI.

In the same year the EU was moving toward final agreement on the AI Act, with fines for violations. The Hiroshima process offered a different route internationally: breadth of participation over binding force, with transparency built up through companies' self-reporting. Japan carried that approach into domestic law; its 2025 AI Promotion Act directs the state to prepare guidelines consistent with the intent of international norms.

Social impact

From the standpoint of users and workers, what the process directly delivered was not rights or remedies but a common yardstick for asking developers to be transparent. Writing watermarking of AI-generated content, protection of intellectual property and personal data, and responses to disinformation into an international code mattered to creators and news organisations who had raised those issues. But whether to report is up to each company, and no one verifies what is reported. As the CSIS analysis suggests, its effect depends on how widely countries and companies take part.

The wider story is on the AI timeline; see also foundation model and large language model.