China vs USA: Who Is Setting the Global AI Safety Standards?
In July 2023, China became the first major economy to issue binding regulations specifically targeting generative AI. Six months later, the European Union passed its comprehensive AI Act. The United States, meanwhile, has relied on executive orders, voluntary commitments, and agency-level guidance—but no comprehensive federal AI law. As AI capabilities accelerate, the question of who sets the rules is becoming as important as who builds the technology. Here's how the US and China are approaching AI safety—and what their divergent strategies mean for the global tech industry.
The US Approach: Innovation First, Regulation Later
The United States has consistently prioritized AI innovation over preemptive regulation. The Biden administration's October 2023 Executive Order on AI was the most comprehensive federal action to date, but it stopped short of creating binding legislation. Under the current administration, the approach has shifted further toward deregulation.
Key Elements of the US Framework
- Voluntary commitments: The White House secured voluntary safety pledges from 15 leading AI companies including OpenAI, Google, Microsoft, Meta, Amazon, and Anthropic. These commitments cover external testing, information sharing, and watermarking—but carry no legal penalties for non-compliance.
- Agency-level regulation: Rather than a single AI law, US regulation is fragmented across agencies. The FTC handles consumer protection, the FDA oversees AI in medical devices, the SEC addresses AI in financial services, and NIST develops technical standards through its AI Risk Management Framework.
- State-level patchwork: California, Colorado, Connecticut, and other states have passed their own AI laws, creating a fragmented compliance landscape. California's SB 1047, the most ambitious state-level AI safety bill, was vetoed in 2024 amid intense industry opposition.
- Export controls as safety policy: The US treats AI chip export restrictions as a form of safety regulation—limiting China's access to advanced semiconductors for AI development. This approach treats AI safety as a geopolitical rather than purely technical challenge.
"The US approach to AI regulation is like building a house room by room, with different architects for each room—and no one has agreed on the blueprint." — AI policy researcher, Stanford HAI
The China Approach: Structured, Centralized, Preemptive
China's AI governance strategy reflects its broader regulatory philosophy: the state sets clear rules, companies comply, and enforcement is centralized. The approach has evolved rapidly since 2021, creating a layered regulatory framework that addresses AI at multiple levels.
Key Elements of China's Framework
- Generative AI Measures (2023): China's Cyberspace Administration (CAC) issued the "Interim Measures for the Management of Generative AI Services," which took effect in August 2023. These require AI providers to ensure content aligns with "socialist core values," prevent discrimination, protect user data, and obtain safety assessments before launching services.
- Algorithm Recommendation Regulation (2022): Before the generative AI rules, China already regulated algorithmic recommendation systems, requiring transparency, user opt-out options, and protection against addictive designs.
- Deep Synthesis Provisions (2023): Specific rules for deepfake and synthetic media technologies require labeling, consent, and content moderation.
- AI Safety Governance Framework (2024): A comprehensive risk classification system categorizes AI systems by risk level, with higher-risk applications facing stricter requirements for testing, documentation, and human oversight.
- Model Registration System: China operates a mandatory registration system for publicly released AI models. Over 200 large language models and AI services had been registered and approved by early 2026, creating a de facto gatekeeping mechanism.
Side-by-Side Comparison
| Dimension | United States | China |
|---|---|---|
| Regulatory Philosophy | Innovation-first, regulate when necessary | Preemptive regulation, structured deployment |
| Legal Framework | No comprehensive federal AI law; executive orders + agency guidance | Multiple binding regulations: Generative AI, Algorithm, Deep Synthesis |
| Enforcement | Fragmented across FTC, FDA, SEC, NIST, state AGs | Centralized: CAC, MIIT, and sector-specific regulators |
| Model Approval | No pre-market approval required | Mandatory safety assessment and registration before public release |
| Content Standards | First Amendment protections; platform-specific content moderation | Must align with "socialist core values"; no content threatening national security or social stability |
| Data Requirements | State-level privacy laws (CCPA, etc.); no federal privacy law | Personal Information Protection Law (PIPL); data must be "legal and legitimate" |
| International Engagement | Bletchley Declaration, AI Safety Summits, bilateral dialogues | Global AI Governance Initiative, UN resolutions, BRI digital cooperation |
| Open Source | Largely unregulated; NTIA report recommended monitoring | Registration required for models above certain capability thresholds |
💡 The Fundamental Difference
The core philosophical divide isn't about whether AI should be safe—both countries agree it should be. The divide is about who decides what "safe" means. In the US model, safety is negotiated between companies, civil society, and regulators. In China's model, the state defines safety unilaterally, and companies comply. Neither model is inherently superior—they reflect different governance traditions and societal values.
Where the US Leads
Technical Standards Development
NIST's AI Risk Management Framework, while voluntary, has become the de facto global reference for AI safety engineering. The framework's detailed taxonomy of AI risks—from data poisoning to model extraction to bias amplification—has influenced regulatory thinking worldwide, including in the EU and Singapore.
The US also leads in AI safety research. Organizations like Anthropic, the Center for AI Safety, and academic labs at Stanford, MIT, and Berkeley are producing foundational research on alignment, interpretability, and robustness. China's AI safety research community is growing but remains smaller and less internationally connected.
Industry Self-Regulation
US companies have developed sophisticated internal safety processes. OpenAI's Preparedness Framework, Anthropic's Responsible Scaling Policy, and Google DeepMind's Frontier Safety Framework represent serious attempts at self-governance. While critics argue these are insufficient without binding enforcement, they represent real operational investment in safety.
International Coalition Building
The US has been more effective at building international coalitions around AI safety. The Bletchley Park AI Safety Summit (2023), the Seoul AI Summit (2024), and the Paris AI Action Summit (2025) all had strong US participation and leadership. The International Network of AI Safety Institutes, initiated by the US and UK, now includes over 20 countries.
Where China Leads
Speed of Regulatory Deployment
China moved from drafting to implementing generative AI regulations in approximately four months—a speed almost unimaginable in Western democracies. This rapid deployment means Chinese AI companies operate with greater regulatory certainty than their American counterparts, who face an uncertain patchwork of state laws and potential federal action.
Comprehensive Coverage
China's regulatory framework covers AI at multiple layers—from the algorithm level to the application layer to the content output. This systematic approach means fewer regulatory gaps. A Chinese AI company knows what rules apply to its recommendation algorithm, its training data, its model outputs, and its user-facing product—all within a single coherent framework.
Enforcement Capacity
China's centralized regulatory structure enables faster enforcement. When the CAC identifies a violation, it can order immediate remediation. In 2024, multiple AI chatbot services were temporarily suspended for failing to meet content safety requirements, and each was back online within weeks after addressing the issues. The US system, by contrast, relies on slower mechanisms like FTC investigations and lawsuits.
Global South Influence
China's AI governance model is gaining traction in developing countries. The Global AI Governance Initiative, launched by China in 2023, has been endorsed by over 50 countries, primarily in Africa, Southeast Asia, and Latin America. For countries building their AI regulatory frameworks from scratch, China's centralized, structured approach can appear more actionable than the US's fragmented model.
Where Neither Leads (Yet)
Despite the competition, several critical AI safety challenges remain unaddressed by both countries:
Frontier Model Evaluation
Neither the US nor China has established a robust, independent evaluation system for frontier AI models. Both countries rely primarily on company self-assessment. The UK and US AI Safety Institutes have conducted limited evaluations, but these remain small-scale and non-binding. Independent, standardized evaluation of models with potentially dangerous capabilities—CBRN knowledge, cyber-offensive capabilities, autonomous replication—remains a gap in both regulatory frameworks.
Open Source Governance
The release of powerful open-source models like Meta's Llama series and China's Qwen and DeepSeek models has created a governance challenge neither country has resolved. Once a model is released openly, it can be fine-tuned to remove safety features. The US has largely avoided restricting open-source AI, while China requires registration for models above certain capability thresholds—but enforcement remains inconsistent.
International Coordination
Despite both countries participating in AI safety summits, there is no binding international agreement on AI safety—no "AI Non-Proliferation Treaty" equivalent. The US and China have different red lines, different risk assessments, and limited trust. This coordination gap means dangerous AI capabilities could emerge without any international mechanism for response.
What This Means for the Global AI Industry
The divergence between US and Chinese AI safety approaches creates both challenges and opportunities:
Compliance Costs for Global Companies
Companies operating in both markets face dual compliance burdens. An AI product that meets US requirements may need significant modification for the Chinese market, and vice versa. This fragmentation increases costs and complexity, particularly for startups and mid-sized companies without large legal teams.
Regulatory Arbitrage
Some companies may choose to develop AI systems in the jurisdiction with more favorable rules. A company developing AI with sensitive applications might choose the US for its lighter regulatory touch, while a company focused on regulated industries might prefer China's clearer rules. This dynamic could shape where different types of AI innovation occur.
The Race to Influence
Both countries are actively exporting their regulatory models. The EU's AI Act, with its risk-based approach, represents a third model that other countries are adopting. The global AI governance landscape is becoming a three-way competition between the US innovation-first model, China's state-led model, and Europe's rights-based model.
"The country that defines AI safety standards will define the AI industry. Standards aren't just about safety—they're about market access, competitive advantage, and technological sovereignty." — International trade lawyer specializing in technology regulation
Conclusion: A Fragmented Future
Neither the US nor China has definitively "won" the race to set global AI safety standards. Instead, the world is heading toward a fragmented regulatory landscape with multiple competing frameworks. The US leads in technical standards and industry self-regulation; China leads in regulatory speed, coverage, and enforcement capacity. Europe offers a third path with its comprehensive, rights-based approach.
For the global AI industry, this fragmentation is both a burden and a reality. Companies will need to navigate multiple regulatory regimes, each with different requirements and enforcement mechanisms. The most successful AI companies will likely be those that build compliance into their systems from the start, rather than treating it as an afterthought.
The ultimate question—whether AI safety standards will converge or diverge further—depends on whether the US and China can find common ground on the most critical risks. On issues like preventing AI-generated CBRN content, preventing autonomous AI systems from causing catastrophic harm, and ensuring AI systems remain under meaningful human control, the two countries' interests align more than their rhetoric suggests. Whether that alignment translates into cooperation remains to be seen.