← Back to Resources

Governing Superintelligence:
A Brief for Policymakers

The case for international AI safety governance: the problem, what is needed, and the precedents that show it is possible.

The Challenge

An unprecedented technology with no governance framework

Artificial superintelligence (a system exceeding human intelligence in every domain) is being actively developed by private corporations operating under minimal regulatory oversight. Leading AI laboratories have stated internally that they expect to reach this threshold within this decade.

Unlike previous technological risks, superintelligence has no existing international governance architecture. There is no treaty, no verification body, no multilateral safety standard, and no mechanism to pause development if warning signs appear. The governance gap is total.

Why It Cannot Wait

The window for effective governance is closing

International governance frameworks require years of negotiation to establish. The Nuclear Non-Proliferation Treaty took a decade of diplomacy after the first nuclear tests. The Montreal Protocol required sustained scientific consensus-building before achieving multilateral adoption.

AI capability development is accelerating. Improvements that once took years now arrive in months. Once a sufficiently capable system exists, it may be too late to impose meaningful constraints. Governance must precede capability, not follow it.

Precedents: We have done this before

Nuclear NPT (1968) International treaty limiting nuclear weapons proliferation. 191 state parties. Includes inspection and verification mechanisms.

Montreal Protocol (1987) Phased out 99% of ozone-depleting substances globally. Called "the most successful environmental treaty in history" by the UN.

Biological Weapons Conv. (1972) Banned development and stockpiling of biological weapons. 183 state parties. Established international norm against bioweapons.

Outer Space Treaty (1967) Prohibited nuclear weapons in space and on celestial bodies. Negotiated between the US and USSR during peak Cold War tension.

Recommended Policy Actions

Mandatory pre-deployment safety evaluations by independent government bodies before any frontier AI system is released, equivalent to clinical trials for pharmaceuticals.
Compute monitoring and licensing for training runs above defined capability thresholds. Compute is the measurable proxy for capability; it can be tracked and regulated.
International AI Safety Agency with inspection authority, modeled on the IAEA, to verify compliance with agreed safety standards across jurisdictions.
Incident reporting requirements for dangerous behaviors observed during training or deployment, currently voluntary and inconsistent across labs.
Domestic AI safety legislation establishing legal liability for foreseeable harms from inadequately tested AI systems, creating appropriate market incentives for caution.
Multilateral negotiations to establish international standards before capability thresholds are reached, not after. Begin now, while the window is open.

The Expert View

This is the scientific consensus, not a fringe position

Over 500 AI scientists, including Nobel laureates Geoffrey Hinton and Yoshua Bengio, and the CEOs of OpenAI and DeepMind, signed a 2023 statement calling extinction-level AI risk "a global priority alongside pandemics and nuclear war."

The UK Government established the world's first AI Safety Institute in 2023. The EU AI Act introduced mandatory risk classification for AI systems. The US Executive Order on AI required safety evaluations for frontier models. The international framework does not yet exist, but the political will is forming. The moment to act is now.

Contact

naotonakada.org

The Nakada Foundation to Save Humanity advocates for international AI safety governance. We provide policy expertise, public education, and advocacy support. Contact us to arrange a briefing for your office or staff.