US States Build a Shared Baseline for Frontier AI Safety

The United States is gradually developing a shared approach to governing the most powerful artificial intelligence systems through parallel action by state and federal authorities. California, New York, and Illinois have advanced legislation focused on frontier-model safety, creating overlapping requirements that could become the foundation of a nationwide policy even before Congress adopts a single federal law. OpenAI describes this process as “reverse federalism”: instead of waiting for Washington to establish rules that states later follow, states adopt compatible safeguards that collectively point the federal government toward a common standard.

The article argues that decisions about frontier AI safety should ultimately be made through democratic institutions rather than left entirely to the companies developing advanced models. A formal national framework would be preferable because it could provide predictable obligations across the country. Until such a framework exists, however, states can approximate one by passing laws built around the same limited set of principles. Aligned state laws could therefore create a de facto national safety baseline while preserving a path toward eventual federal regulation.

Consistency is presented as essential both for safety and for American technological leadership. A fragmented collection of incompatible state rules could slow development, complicate enforcement, confuse consumers, and force companies to spend resources on compliance work that does not directly improve safety. The burden would be especially significant for startups and smaller developers. OpenAI also argues that excessive fragmentation could weaken efforts to build a democratic AI ecosystem capable of supplying governments, critical infrastructure operators, allies, and trusted partners with tools to defend against malicious uses of AI, including cyberattacks.

The proposed state-level baseline has two central components. Developers of frontier models should maintain a documented safety framework, conduct risk assessments, and publicly disclose both the assessments and their results. Their claims and procedures should also be subject to independent, objective audits that provide governance and accountability. California established the central disclosure structure, New York demonstrated that a similar approach could be adopted in another jurisdiction, and Illinois added a requirement for independent verification of important disclosures. Together, the three states are described as embedding democratic oversight into the deployment of frontier AI.

The article acknowledges that legislation frequently accumulates additional provisions as lawmakers negotiate the votes required for passage. Nevertheless, it urges states to remain disciplined and concentrate on the core safety and accountability measures. Expanding state laws into unrelated or overly broad areas could produce policy creep and recreate the very regulatory patchwork that alignment is meant to prevent. Such complexity would be difficult for regulators to administer and could divert engineering and financial resources away from practical safety work.

The author also draws a boundary between appropriate state oversight and responsibilities that should remain federal. States should not be expected to assess major national-security threats or effectively make security decisions for the entire country. They also may not have the classified access, specialist expertise, technical resources, or sustained capacity required to evaluate the most advanced systems. Those tasks are better handled by federal experts who can work directly with AI developers and coordinate policy across security agencies and international partners.

At the federal level, the Trump Administration is working with technical and national-security specialists on a framework for US government testing of the most capable AI models for cybersecurity risks and capabilities. The planned framework is expected to define testing standards, schedules, and processes. OpenAI says it is participating in discussions with the administration, other AI companies, business organizations, and additional stakeholders as that system is developed.

Cybersecurity evaluations are used as an example of why a stable national process is necessary. Advanced models are already being tested while the federal framework remains unfinished, which the article describes as understandable given the pace of development. Experience from these early evaluations has nevertheless shown the need for a consistent, repeatable method. A coherent national system would make model testing more reliable and could later support an international safety framework based on democratic values.

The broader ambition is for the United States to convert domestic alignment into global leadership. A national standard could provide the basis for international rules governing safe AI deployment while helping democratic countries retain access to defensive capabilities. Achieving that outcome, the article argues, requires sustained coordination among states, the federal government, AI laboratories, industry, allies, and international institutions. Symbolic legislation or politically performative measures would not be sufficient; durable safety policy depends on compatible rules, clearly assigned responsibilities, and institutions capable of applying them in practice.