US Rejects Global AI Standards from OpenAI and Anthropic
US Rejects Pleas from OpenAI and Anthropic for Global AI Standards
The debate over artificial intelligence governance has shifted from academic discourse into high-stakes global statecraft. As foundational model capabilities accelerate, leading artificial intelligence developers have lobbied governments to establish centralized international guardrails. However, the United States government has formally declined calls from industry leaders—most notably OpenAI and Anthropic—to implement binding global AI safety standards Source 1, Source 3.
This decision marks a critical turning point in international technology policy. The US choice to prioritize domestic governance and private-sector agility over multilateral treaties directly affects how frontier AI models are built, audited, and deployed globally.
1. Overview of the US Policy Decision
┌────────────────────────────────────────┐
│ OpenAI & Anthropic Global AI Appeals │
│ • Multilateral compute thresholds │
│ • Treaty-level safety verification │
│ • Centralized international body │
└───────────────────┬────────────────────┘
│
▼
┌────────────────────────────────────────┐
│ US Executive Policy Stance │
│ [ REJECTION OF BINDING FRAMEWORK ] │
└───────────────────┬────────────────────┘
│
┌────────────────────────────────┼────────────────────────────────┐
▼ ▼ ▼
┌─────────────────┐ ┌──────────────────┐ ┌─────────────────┐
│ National Tech │ │ Geopolitical │ │ Agile Domestic │
│ Innovation Pace │ │ Competitiveness │ │ Risk Frameworks │
└─────────────────┘ └──────────────────┘ └─────────────────┘
Key Details of the Rejection
The United States rejected proposals from top AI development firms seeking international treaties to standardize safety, alignment, and evaluation protocols for frontier AI systems Source 1, Source 3.
Instead of binding multilateral agreements, federal policymakers are maintaining a national, decentralized regulatory strategy. The decision follows extensive appeals from leading frontier AI labs that urged the executive branch and international coalitions to coordinate universal compute governance and deployment restrictions. Federal officials determined that formal international treaties would introduce administrative bottlenecks, hinder domestic research velocity, and limit national strategic autonomy in software engineering.
The Push for Centralized Global Oversight
The rejected proposals originated from concerns raised by OpenAI and Anthropic regarding catastrophic risks from unaligned foundation models. Both organizations argued that frontier models—systems trained on unprecedented scale that exhibit novel autonomous behaviors—pose systemic risks that exceed the jurisdiction of single-nation regulators.
Their advocacy focused on three primary areas:
- Frontier Model Containment: Establishing technical barriers against the weaponization of autonomous agents in cybersecurity, biotechnology, and critical infrastructure.
- Catastrophic Risk Mitigation: Mandating rigorous red-teaming and failsafes before systems achieve advanced autonomy or recursive self-improvement.
- Unified Alignment Metrics: Establishing universally recognized technical benchmarks to quantify alignment, interpretability, and output safety across jurisdictions.
2. Proposals Put Forward by OpenAI and Anthropic
Frontier Lab Governance Blueprint:
┌─────────────────────────┐ ┌─────────────────────────┐ ┌─────────────────────────┐
│ Compute Registration │ ──> │ Third-Party Red-Teaming │ ──> │ Supranational Oversight │
│ (>10^26 Floating Ops) │ │ (Pre-Deployment Audits) │ │ (IAEA-Style Agency) │
└─────────────────────────┘ └─────────────────────────┘ └─────────────────────────┘
Core Demands of the AI Frontier Labs
OpenAI and Anthropic outlined a comprehensive regulatory framework aimed at governing the top tier of AI capabilities. Their strategy centered on institutional controls designed to manage dangerous models without restricting lower-risk software.
| Proposal Area | Proposed Implementation Mechanism | Target Risk Area |
|---|---|---|
| Compute Thresholds | Mandatory tracking of training clusters exceeding designated FLOP counts (e.g., $10^{26}$ floating-point operations). | Proliferation of unmonitored frontier foundation models. |
| Pre-Deployment Audits | Mandatory third-party red-teaming and safety verification prior to public weights release or API launch. | Zero-day vulnerability exploitation, cyberattacks, and chemical/biological risk vector generation. |
| International Oversight | Creation of a specialized global agency similar to the International Atomic Energy Agency (IAEA) or civil aviation regulators. | Unilateral, unsafe AI model development by state and non-state actors. |
| Safety Governance Plans | Standardized Responsible Scaling Policies (RSP) tied directly to hardware scaling benchmarks. | Unchecked runaway model training without safety validation. |
These proposals emphasized that digital compute clusters, advanced semiconductors, and hyper-scale data centers represent physical choke points suitable for international inspection regimes.
Corporate Motivations Behind Regulatory Advocacy
The push for global regulation by leading labs is driven by risk management and commercial stability:
┌──────────────────────────────────────────────┐
│ Corporate Motivations for Global Regulation │
└──────────────────────┬───────────────────────┘
│
┌────────────────────────────────┴────────────────────────────────┐
▼ ▼
┌─────────────────────────────────┐ ┌──────────────────────────────────┐
│ Liability & Standardization │ │ Non-State & Hostile Parity │
├─────────────────────────────────┤ ├──────────────────────────────────┤
│ • Clear statutory boundaries │ │ • Global compute barriers │
│ • Protection from civil torts │ │ • Hardware access verification │
│ • Uniform cross-border audits │ │ • Neutralizing rogue deployment │
└─────────────────────────────────┘ └──────────────────────────────────┘
- Liability Distribution and Compliance Predictability: Clear, unified legal standards provide market protection. Operating across dozens of contradictory state-level and national regulatory frameworks introduces legal liabilities. A single international standard creates predictable compliance pathways and protects compliant developers from broad civil litigation.
- Protection Against Hostile and Unregulated Deployments: Voluntary domestic adherence to expensive safety evaluations places regulated firms at a competitive disadvantage if rival entities or foreign adversaries release high-risk systems without safeguards. International governance was intended to prevent unconstrained developers from capturing market share with high-risk models.
3. Strategic Drivers Behind the US Rejection
US Policy Rationale
│
┌─────────────────────────────────┼─────────────────────────────────┐
▼ ▼ ▼
┌─────────────────────────┐ ┌─────────────────────────┐ ┌─────────────────────────┐
│ Technical Acceleration │ │ Geopolitical Asymmetry │ │ Agile Governance Models │
├─────────────────────────┤ ├─────────────────────────┤ ├─────────────────────────┤
│ Rigid treaties create │ │ Non-aligned actors │ │ NIST frameworks provide │
│ administrative drag on │ │ bypass treaties; US │ │ iterative, domestic │
│ domestic compute hubs. │ │ compliance limits edge. │ │ risk mitigation tools. │
└─────────────────────────┘ └─────────────────────────┘ └─────────────────────────┘
Technological Primacy and Innovation Velocity
The primary driver behind the US rejection is the imperative to maintain commercial and technical leadership in generative AI and high-performance computing.
- Treaty Drag: Negotiating, passing, and maintaining multilateral agreements requires extensive bureaucratic coordination. The multi-year timelines required to update international treaties conflict with the rapid pace of frontier model development.
- Venture and Capital Dynamics: Strict operational requirements risk slowing capital deployment across the broader software and hardware stack, including data centers, semiconductor fabrication, and cloud infrastructure.
- Open-Source Ecosystem: Broad international restrictions on model weights or compute clusters could undermine open-source software, a major driver of US developer adoption and technological influence worldwide.
Geopolitical Competition and National Security
AI compute is recognized as a foundational component of national security, intelligence analysis, and electronic warfare.
- Enforcement Deficits: Multilateral treaties depend on verify-and-comply mechanisms. The US government identified significant structural challenges in verifying model safety inside adversarial or non-aligned jurisdictions.
- Strategic Asymmetry: Binding the US technology ecosystem to treaty inspections without reliable means to enforce equivalent compliance abroad creates strategic vulnerabilities.
- Sovereign Standards Over Global Consensus: The federal government prefers domestic control over critical technologies over international bodies that might be influenced by geopolitical competitors.
Reliance on Voluntary Commitments and Domestic Frameworks
Rather than creating international treaty obligations, the US approach uses flexible domestic structures designed to adapt alongside emerging models:
US Domestic Policy Architecture
│
┌─────────────────────────────┴─────────────────────────────┐
▼ ▼
┌──────────────────────────────────────┐ ┌────────────────────────────────────┐
│ Non-Binding Risk Standards (NIST) │ │ US Artificial Intelligence Safety │
│ Iterative AI Risk Management Engine │ │ Institute (AISI) Research Channel │
└──────────────────────────────────────┘ └────────────────────────────────────┘
- NIST AI Risk Management Framework (AI RMF 1.0): Provides non-punitive, adaptable technical guidelines for organizations to identify, measure, and manage algorithmic bias, safety issues, and security vulnerabilities.
- Executive Orders and Domestic Safety Institutes: Uses national mechanisms—such as the US Artificial Intelligence Safety Institute (US AISI)—to evaluate models via bilateral agreements with developers rather than rigid legal mandates.
- Voluntary Private Sector Pledges: Relies on voluntary public commitments from leading technology firms regarding watermarking, red-teaming, and security audits to avoid legislative gridlock.
4. Global Market and Regulatory Fragmentation
The refusal of the US to adopt centralized global standards accelerates the fragmentation of the global technology market into distinct regulatory environments.
Global AI Governance Divergence
│
┌───────────────────────────────┴───────────────────────────────┐
▼ ▼
┌──────────────────────────────────────────┐ ┌──────────────────────────────────────────┐
│ European Union (EU AI Act) │ │ United States Model │
├──────────────────────────────────────────┤ ├──────────────────────────────────────────┤
│ • Pre-market certification requirements │ │ • Non-binding, voluntary risk management │
│ • Strict fines for non-compliance │ │ • Market-driven innovation cycles │
│ • Categorized systemic risk obligations │ │ • Sectoral agency enforcement │
└──────────────────────────────────────────┘ └──────────────────────────────────────────┘
Divergence Between the US and Global Frameworks
The international regulatory landscape is divided into three distinct operational methodologies:
┌───────────────────────────────┐
│ United States Model │ ──> Market-driven, voluntary safety benchmarks, decentralized enforcement.
└───────────────────────────────┘
┌───────────────────────────────┐
│ European Union Model │ ──> Prescriptive, tiered risk categories, centralized market-entry barriers.
└───────────────────────────────┘
┌───────────────────────────────┐
│ Multilateral Initiatives │ ──> Soft-law diplomatic communiqués (e.g., G7 Hiroshima AI Process).
└───────────────────────────────┘
- The European Union (EU AI Act): Employs an enforceable, risk-tiered system. It introduces administrative fines for non-compliance, bans specific use cases (e.g., real-time biometric social scoring), and mandates detailed transparency obligations for general-purpose AI (GPAI) systems with high systemic impact.
- The United States: Relies on market-driven innovation supported by sectoral agency enforcement (such as the FTC on consumer deception, SEC on disclosures, and DOJ on antitrust) and voluntary safety protocols.
- Multilateral Working Groups: Forums like the G7 Hiroshima Process, the OECD AI Principles, and United Nations advisory bodies produce non-binding guidance that lacks enforcement mechanisms without state-level ratification.
Impact on Multinational Enterprise Operations
This regulatory divide creates significant operational challenges for enterprise technology providers:
- Cross-Border Compliance Costs: Multinational software providers must maintain separate product architectures to operate legally across jurisdictions. A system deployed freely in the US market may require deep algorithmic auditing, system documentation, and data-lineage certification before release in the European single market.
- Jurisdiction Shopping and Regulatory Arbitrage: Startups and independent research groups can relocate compute workloads, incorporated entities, or model deployments to countries with minimal safety oversight or intellectual property restrictions.
- Hardware and IP Isolation: Export controls on high-end graphic processing units (GPUs) and extreme ultraviolet lithography (EUV) systems widen the operational divide between regulated enterprise environments and offshore infrastructure.
5. Long-Term Implications for AI Safety and Deployment
Long-Term Trajectory of AI Safety Governance
│
┌─────────────────────────────┴─────────────────────────────┐
▼ ▼
┌──────────────────────────────────────┐ ┌────────────────────────────────────┐
│ Systemic Safety Challenges │ │ Alternative Coordination Pathways │
├──────────────────────────────────────┤ ├────────────────────────────────────┤
│ • Uneven global safety testing │ │ • Bilateral technical compacts │
│ • High risk of regulatory arbitrage │ │ • Standardized API-level metrics │
│ • Unmonitored frontier deployments │ │ • Industry-led safety alliances │
└──────────────────────────────────────┘ └────────────────────────────────────┘
Risks of Uncoordinated Frontier Model Deployment
The absence of a centralized international safety framework introduces several systemic risks:
- Inconsistent Safety Auditing: Without universal benchmarks, the thoroughness of pre-deployment audits depends on voluntary corporate policies or fragmented national requirements.
- Vulnerability to Systemic Shocks: Automated cyber defense evasion, automated zero-day generation, or unmonitored chemical/biological synthesis designs present cross-border challenges that domestic regulations cannot individually contain.
- Coordination Failures During Accidents: If a frontier model experiences an alignment failure, the lack of secure, cross-border incident-reporting protocols could delay effective technical mitigation.
Alternative Pathways for Global AI Coordination
Because a centralized international treaty remains off the table, international AI coordination is evolving through alternative, decentralized channels:
Decentralized Governance Framework:
┌─────────────────────────────┐ ┌─────────────────────────────┐ ┌─────────────────────────────┐
│ Bilateral AISI Pacts │ ──> │ Compute Hardware Controls │ ──> │ Standardized Evaluations │
│ (US-UK Safety Collaborations)│ │ (Semiconductor Supply Audits)│ │ (Benchmarking & Open Eval) │
└─────────────────────────────┘ └─────────────────────────────┘ └─────────────────────────────┘
- Bilateral Technical Pacts: Direct agreements between targeted safety organizations, such as the formal collaboration between the US and UK AI Safety Institutes on pre-deployment model evaluations.
- Hardware-Layer Export Controls: Using specialized supply chains—such as advanced foundries and semiconductor manufacturing tooling—to enforce practical capability limits globally without comprehensive software treaties.
- Standardized Developer Evaluations: Grassroots industry adoption of unified evaluation suites (such as EleutherAI benchmarks, inspect-ai, or METR testing protocols), establishing de facto industry standards from the engineering level up rather than top-down political directives.
Frequently Asked Questions (FAQ)
What was the core proposal from OpenAI and Anthropic?
OpenAI and Anthropic proposed establishing binding, international regulatory standards to govern the development and deployment of frontier artificial intelligence models Source 1, Source 3. Their plans included multilateral compute thresholds, mandatory third-party pre-deployment red-teaming, and the creation of an international safety monitoring body modeled after civil aviation or nuclear energy regulators.
Why did the United States reject global AI standards?
The United States declined the proposals to protect domestic technological innovation, maintain agility in frontier research, and protect its technological edge Source 1, Source 3. US policymakers determined that formal international treaties would introduce administrative bottlenecks, prove difficult to verify among geopolitical adversaries, and restrict domestic industrial capabilities.
How does the US approach to AI safety differ from the European Union?
The US approach focuses on non-binding, voluntary risk management frameworks (such as the NIST AI RMF), bilateral research partnerships through the US AI Safety Institute, and standard market-driven protections. In contrast, the European Union enforces the EU AI Act, a comprehensive, risk-tiered law with mandatory technical compliance, pre-market certifications, and administrative penalties for non-compliance.
What are the main risks of lacking unified global AI standards?
The primary risks include inconsistent safety evaluations across different regions, regulatory arbitrage by developers seeking to avoid safety testing, increased compliance costs for multinational companies, and the lack of coordinated emergency response systems if a frontier model causes catastrophic cross-border harm.
Will the US participate in any international AI governance initiatives?
Yes. The United States continues to engage in non-binding multilateral discussions, including the G7 Hiroshima AI Process, OECD working groups, and bilateral evaluation agreements with international partners like the United Kingdom. However, it avoids signing binding, treaty-level agreements that place operational limits on domestic AI development.
