Why Global AI Regulation Needs a New Institution—and What It Might Look Like

Introduction

In June 2026, Shane Legg, co‑founder of Google DeepMind, warned that the world builds rockets without a single fuel standard, highlighting the lack of a unified global AI regulatory framework. The EU’s AI Act, the United States’ patchwork of federal guidance and state laws, and China’s scenario‑specific regulations create a fragmented governance vacuum that raises real safety risks.

Current Situation: One Model, Many Rules

More than 60 jurisdictions have AI‑related laws, but fewer than 15% are compatible. A San‑Francisco AI firm must satisfy at least three distinct compliance regimes to launch a large‑model product in the EU, Japan, and Brazil, covering transparency reports, risk‑level grading, and complaint‑response timelines.

Concrete example: Meta’s Llama 4 launch in Brazil required a Portuguese‑language explainability report that differed from the EU AI Act format, forcing the compliance team to produce two separate document sets over three months.

OECD (2026) estimates that multinational AI firms now spend 3.2% of annual revenue on multi‑jurisdiction compliance, up from 0.8% two years earlier.

Divergence Among the Three Major Economies

EU follows a “risk‑tiered” approach, classifying AI systems into four risk levels; high‑risk models need third‑party conformity assessment.

United States relies on industry self‑regulation plus targeted interventions; federal guidance (NIST AI RMF 2.0) is voluntary, while states enact their own rules (e.g., California’s SB 1047, Texas restrictions on energy‑sector AI).

China adopts “scenario‑specific legislation,” issuing rapid but often uncoordinated rules for each emerging AI application.

Technical Challenges of Cross‑Border Governance

1. Inconsistent model evaluation standards : EU robustness tests, NIST metrics, and China’s TC260 standards differ, leading to cases like a European bank that had to redo a fraud‑detection model test using EU data sets after the US supplier’s NIST report was rejected.

2. Data sovereignty vs. training pipelines : India’s 2025 Personal Data Protection law, EU GDPR, and Middle‑East localisation mandates make a single global training pipeline practically impossible.

3. Incident response coordination : A hypothetical AI‑driven medical diagnosis system deployed in Singapore serving Southeast Asian users could cause a misdiagnosis affecting patients in Thailand, developers in the US, and data stored in Japan, with no clear authority to lead the investigation.

Why Existing International Bodies Cannot Fill the Gap

Voluntary frameworks like the OECD AI Principles lack enforcement; the newly created UN AI advisory body has only 23 staff, half of whom have technical AI expertise; and G7/G20 initiatives avoid contentious topics such as military and surveillance AI.

The article argues that a model similar to the International Telecommunication Union (ITU) or International Civil Aviation Organization (ICAO) is needed—an institution that creates enforceable technical standards rather than merely discussing principles.

Proposed Architecture for a Global AI Standards Institution

The suggested three‑layer design includes:

Unified technical standards (e.g., baseline safety assessment methods and reporting formats) with implementation left to national regulators, mirroring ICAO’s model.

Mutual recognition of assessments so that certification in one jurisdiction is accepted elsewhere, akin to International Financial Reporting Standards (IFRS).

An independent incident coordination center, similar to the WHO’s International Health Regulations mechanism, capable of launching a 48‑hour joint investigation.

Implementation Path and Real‑World Obstacles

Geopolitical resistance : Major powers such as the US and China are unlikely to cede control over AI standards that affect strategic industries.

Speed of technological change : AI advances monthly, whereas traditional standards bodies take years; the new institution must adopt a faster, iterative standard‑setting process.

Resource and legitimacy constraints : Comparable agencies like the IAEA operate with budgets of $650 million and thousands of staff; securing similar funding and talent for an AI standards body is a major challenge.

A three‑step rollout is proposed:

Start with “minimum consensus” on low‑controversy, high‑utility topics such as model safety assessment methods, high‑risk AI definitions, and incident‑report formats.

Leverage existing bilateral agreements (e.g., EU‑Japan AI governance memorandum) to pilot “one assessment, multiple recognitions.”

Expand into a multilateral mechanism once a dense network of mutual recognitions creates demand for a central coordinating entity.

Conclusion

Fragmented AI regulation will not resolve itself; as AI systems become more capable and pervasive, cross‑border governance gaps will widen. A dedicated global AI standards institution would not be a cure‑all, but it could provide a common technical language that aligns regulators, developers, and users, reducing compliance costs and improving safety—much like the historic creation of ICAO or the IAEA during periods of rapid technological risk.