The G7 Hiroshima Process: soft law at ChatGPT speed

Lesson 4 of 5 in Global Soft Law: OECD, UNESCO, and the G7 Hiroshima Process.

ChatGPT launched in November 2022. Within six months, at the May 2023 Hiroshima summit under Japan’s presidency, G7 leaders ordered up a governance response; by October 2023 it existed. Compare the clocks: the EU AI Act took roughly five years from proposal to application; the Hiroshima Process produced its instruments in about five months. That speed is the whole argument for soft law — and Japan, whose own domestic approach favours agile, non-binding governance, was the natural author.

The October 2023 package has two layers. The International Guiding Principles — eleven of them — address organizations developing advanced AI systems: identify and mitigate risks across the lifecycle, monitor for misuse after deployment, publish transparency reports, share information with peers and governments, invest in security controls, deploy content authentication and provenance mechanisms such as watermarking, prioritize research on societal risks, and develop standards. The companion International Code of Conduct turns those principles into a checklist a company can adopt — and be measured against. Notice the audience shift: the OECD speaks to governments, UNESCO to states and societies, but Hiroshima speaks directly to the dozen-odd firms building frontier models.

Why does a voluntary report matter? Three reasons professionals care. First, comparability: every filer answers the same questions, so a researcher, journalist, or regulator can line up what each company claims about its safety practices — the first common disclosure channel the industry has had. Second, commitment escalation: a public filing is evidence. A company that files "we red-team all frontier releases" and then demonstrably does not has handed enforcement agencies a deception theory — recall from the foundations module how the US FTC treats broken voluntary commitments. Third, template effect: when binding law arrives (the EU’s GPAI obligations, for instance), regulators reach for the disclosure formats that already exist.

The process also grew a diplomatic layer: the Hiroshima Process Friends Group, launched in 2024 and open beyond the G7, gathered roughly fifty countries endorsing the principles — the standard soft-law expansion play you saw with OECD adherence.

Key terms: Hiroshima AI Process, code of conduct, transparency report, watermarking (AI content)

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.