A proposed international regulatory framework for artificial intelligence should focus on institutions and governance, not just technical controls, recognizing regulation as political, economic, and legal.

Enforcement relies on incentives and mutual dependence rather than force, granting compliant states benefits such as market access, partnerships, and resources, while imposing restrictions on non-compliant jurisdictions to influence behavior.

Transparency is a core pillar, requiring detailed records on data sources, model architecture, computational resources, and limitations; regulators must have inspection rights when needed, even if some information is confidential.

Historical analogies from aviation, telecommunications, maritime navigation, nuclear energy, and civil aviation show that global standards can reduce risk while preserving innovation; AI requires a similar governance approach.

Military applications demand separate treatment, with a convention prohibiting autonomous weapons that can select and engage targets without meaningful human oversight, while allowing non-lethal military AI uses under human responsibility.

Governance should be bicameral to balance power: one chamber equal for states, and another based on factors like economic contribution, computational capacity, and research output; major decisions require consensus or cross-chamber approval.

The overarching aim is to balance safety and innovation, avoiding excessive caution that stifles development or unchecked competition that neglects safety, framing AI governance as a constitutional challenge about power distribution and accountability.

Economic concentration and competition concerns require cooperation among international competition authorities, promoting interoperability, open standards, and data portability to prevent monopolistic control of frontier AI.

AI regulation should be risk-based with a four-tier system—low, medium, high risk, and frontier/major systems—each with appropriate requirements such as transparency, auditing, licensing, and international oversight.

Control of computational resources is a regulatory focus, emphasizing large data centers and facilities over individual researchers, potentially via international licensing for facilities above certain thresholds.

Liability should rest with humans: developers for design defects, deployers for operational misuse, corporate officers for negligence or recklessness; AI systems should not be granted legal personhood.

An International Artificial Intelligence Agency is proposed, modeled after the IAEA, with universal membership, standards-setting, regulator accreditation, inspections of frontier facilities, and incident investigations, while preserving domestic sovereignty and minimum international standards.