OpenAI urges US to set global safety standards for AI, proposing a US-led network of 10 nations to define incident reporting, evaluation, and rules for self-improving models.

OpenAI urges US to set global safety standards for AI

OpenAI urges US to set global safety standards for AI, calling on Washington to lead an international coalition to define technical rules for frontier models, incident reporting, and self-improving systems. In a blog post published on September 21, 2026, the company said shared standards are now essential to keep AI development safe as capabilities advance.

The proposal marks a shift from OpenAI’s earlier resistance to binding government rules. The company now wants the US to coordinate with other nations to create common benchmarks, reporting requirements, and governance frameworks for the most capable AI systems.

What OpenAI is proposing

OpenAI’s plan centres on three main pillars:

• Incident reporting standards: Clear criteria for which AI-related incidents must be reported, how they should be classified, and what information labs and governments should share.
• Evaluation and safety benchmarks: Common tests and risk-assessment methods for frontier models, including systems that can improve themselves through automated research.
• Compute and cooperation rules: Guidelines for how countries collaborate on AI advancement and how access to high-end computing resources is managed to reduce misuse risks.

The company argues that without coordinated standards, the world risks a “fragmented, uneven, and conflict-ridden system” in which different regions impose incompatible rules and safety gaps emerge.

A US-led network of AI safety institutes

Rather than creating a new global body from scratch, OpenAI suggests building on existing national AI safety organisations. It proposes a network of about 10 countries with established AI safety institutes, including:

• The United States (via the Center for AI Standards and Innovation, CAISI, within the Commerce Department).
• The United Kingdom (AI Safety Institute).
• France, Singapore, South Korea, and others that have launched similar bodies.

This network would:

• Develop shared technical standards for frontier AI.
• Align national regulations so they complement rather than conflict with each other.
• Coordinate on risk assessment, incident response, and best practices for labs.

OpenAI says the US is “well positioned to lead because its AI industry is at the technical frontier,” and that acting now will determine whether the US shapes the global framework or is forced to adapt to one designed by others.

Focus on self-improving AI

A key part of the proposal concerns recursive self-improvement (RSI) and automated AI research, where models can modify or train other models with minimal human intervention. OpenAI states:

“Recursive self improvement, or RSI, should not be pursued ‘unless and until it can be done safely.’”

The company wants international standards to:

• Define what counts as RSI or automated research.
• Set safety thresholds before such systems are deployed.
• Require specific evaluations and monitoring for models capable of self-improvement.

This reflects growing concern that self-improving systems could accelerate capability gains faster than safety measures can adapt.

Timing and political context

The proposal comes as:

• CEO Sam Altman prepares to brief the United Nations General Assembly on AI risks and governance.
• Governments worldwide debate how to regulate frontier AI, with the EU AI Act already requiring certain incident reporting since 2025.
• Industry leaders such as Anthropic’s Dario Amodei, OpenAI’s Sam Altman, and xAI’s Elon Musk recently backed calls to “pace the frontier” of AI development.

By calling for US-led global standards, OpenAI is positioning itself as a pro-regulation voice that prefers coordinated international rules over a patchwork of national laws.

What happens next

OpenAI’s proposal is a policy recommendation, not a law. Next steps could include:

• Discussions with the US Commerce Department, CAISI, and other agencies.
• Engagement with allied governments and their AI safety institutes.
• Potential incorporation of OpenAI’s ideas into upcoming US or international AI governance initiatives.

The company has not specified a timeline for when standards should be finalised or how compliance would be enforced.

Summary: OpenAI urges US to set global safety standards for AI, proposing a US-led network of about 10 nations to define incident reporting, evaluation benchmarks, and rules for self-improving systems. The plan builds on existing AI safety institutes and marks a shift from OpenAI’s earlier opposition to binding government rules.