Acceleration of AI Governance: Is Global Policy Heading Toward Fragmentation?

In recent years, major countries around the world have been accelerating the development and implementation of AI governance policies. China has officially begun trial implementation of its AI Ethics Management Rules, the US White House is considering pre-approval systems for advanced AI models, the EU has introduced transparency enforcement guidelines under its AI Act, and Canada’s Office of the Privacy Commissioner has ruled that OpenAI’s training data usage was illegal. Additionally, the newly appointed Pope Leo XIV has openly expressed support for AI regulation. These developments indicate that global AI governance has shifted from debating “whether to regulate” to competing on “how to regulate.”

This shift is no coincidence. As technological capability doubles approximately every four months, the cost of governance gaps is also rising exponentially. Below, we outline the four distinct approaches to AI governance currently being pursued.

Global AI Governance Moves to Implementation Stage

AI governance has now officially shifted from the question of “whether to regulate” to the practical challenge of “how to regulate.” In response to the rapid pace of technological evolution, countries are exploring their own unique pathways to governance.

China: Institutionalizing Comprehensive AI Ethics Review

China’s “Administrative Measures for the Ethical Management of Artificial Intelligence Technology (Trial)” has entered the implementation stage, enforcing comprehensive reviews for all AI research and development organizations. Unlike the EU’s AI Act, which takes a tiered approach based on risk levels, China does not categorize models by size. The review criteria encompass both technical safety and social public interest, adopting a dual-track system of “self-assessment + external review” to ensure flexibility. Multinational companies operating in China must comply with this independent regulatory framework, separate from those in Western countries.

United States: Ongoing Debate Over Advanced AI Pre-Approval Policies

The refusal by Anthropic to release its advanced AI model, Claude Mythos, has reignited debates in Washington about pre-approval systems for advanced AI. Initially, the White House considered an executive order for pre-approval similar to FDA drug approvals but later softened its stance to avoid stifling innovation through bureaucracy. Supporters advocate for licensing tests for models exceeding certain thresholds, while opponents argue that AI is a dynamic general-purpose technology, making the FDA framework inapplicable. Critics also highlight risks of politicizing technical decisions and the rapid obsolescence of capability thresholds as technology evolves. Alternatives like post hoc accountability, mandatory disclosures, and AI insurance are gaining traction. The US AI policy oscillates sharply between deregulation and stricter regulation, with debates on licensing systems coming full circle.

EU: Exporting Global Rules, Moving from Legislation to Execution

The EU’s AI Act, set to be enacted in 2024, has recently advanced three key initiatives: a draft transparency enforcement guideline, the launch of the 28DIGITAL project to offset compliance costs for small and medium-sized enterprises, and a framework document aimed at exporting global AI standards. While the EU does not produce advanced AI models, it is focused on defining usage rules for such technologies. At a time when the US is still debating “whether to regulate,” the EU is deeply engaged in “how to enforce regulation.” Whether the “Brussels Effect” seen with GDPR can be replicated in the AI domain depends on the feasibility of these execution frameworks.

Canada: First Global Ruling on AI Training Data Violation

Canada’s Office of the Privacy Commissioner has ruled that OpenAI’s use of training data for ChatGPT violated the country’s privacy laws, citing the extensive use of publicly available data and the inability of citizens to foresee their data being used for AI training. This ruling is seen as a dangerous precedent, as restricting the use of public data could hinder AI development and place Canada at a disadvantage in the global AI race. The ripple effects of this decision are being closely watched, with European data protection agencies potentially following suit and Asia-Pacific countries also paying attention.

Fragmentation in AI Governance: A Central Challenge

China, the US, Europe, and Canada have each developed self-contained governance frameworks: institutionalized ethics reviews, innovation-prioritized regulatory delays, rule-driven implementation, and privacy-focused limitations on data use. While no single approach is definitively right or wrong, a key contradiction arises from the cross-border nature of AI models versus the fragmented governance rules. This fragmented governance presents challenges even more significant than those posed by GDPR. While data can be stored locally, intelligence cannot be confined locally. Over the next 12 months, the critical question will not be which approach prevails but whether these diverse approaches can establish minimum mutual recognition mechanisms. Without such coordination, fragmentation itself could become the greatest obstacle to AI development.