OpenAI wants state AI safety laws to become the national standard

OpenAI published a policy essay on July 15, written by Chief Global Affairs Officer Chris Lehane, that makes a specific argument: in the absence of federal AI legislation, states are converging on the same core safety requirements, and that convergence is itself becoming the national standard. OpenAI calls this "reverse federalism" — instead of Congress setting a floor that states build on, states are setting a floor Congress can ratify.

This is a lobbying position, not a product or research release, and it's worth reading it as one. OpenAI has an obvious stake in which version of AI regulation wins. But the mechanics it describes are real and worth understanding if you build with frontier models, because they describe the compliance surface heading your way regardless of how the essay's advocacy lands.

The three requirements doing the work

The essay points to California, New York, and Illinois as the states that have converged on the same core elements: a documented safety framework with risk assessments for frontier models and public disclosure of those assessments, mandatory reporting of serious safety incidents, and governance through independent, objective audits. In OpenAI's framing, California built the disclosure baseline, New York proved it travels across jurisdictions, and Illinois added independent verification on top. Stack those three and you get something close to a full external assurance regime — assessed, disclosed, reported, and audited — for any lab whose models clear a frontier-capability bar.

Federal and global are still catching up

At the federal level, the essay points to the Trump Administration's work on a testing framework for evaluating the most capable models on cyber risk, which OpenAI says it hopes lands by early August, built around the Center for AI Standards and Innovation (CAISI). It also name-checks Reps. Jay Obernolte and Lori Trahan among the lawmakers who've put forward federal framework proposals. Globally, the essay ties this back to a G7-adjacent conversation with Brazil, Egypt, India, Kenya, and Korea, plus Sam Altman's Financial Times op-ed proposing a US-led international forum for AI standards. The throughline OpenAI wants readers to take away: state → federal → global, each layer building on the one below it.

What this means if you're building

Strip the advocacy framing and the three-part requirement — safety framework, incident reporting, independent audit — is close to what disciplined engineering teams already do internally: model cards, red-team results, eval suites, and change logs. What's new isn't the practice, it's making it external, disclosed, and auditable by a third party. If your organization is anywhere near frontier capability, or works with a lab that is, this is the shape of the paperwork to plan for: a standing safety framework document, an incident-reporting channel with defined severity thresholds, and an audit trail that survives someone outside your company reading it.

My take

The reverse-federalism logic is sound as description — three states converging on the same three pillars is a real pattern, not a rhetorical stretch. But it's worth noticing who's writing the standard it's converging toward: the frontier labs it will govern are also the ones shaping its content in statehouses and in Washington. A safety framework that mirrors what a well-resourced lab already runs internally is a reasonable floor for that lab and a nontrivial lift for a smaller one — the essay itself concedes this risk when it flags diverted "developer resources... especially at start-ups and small companies." The federal testing framework OpenAI is counting on to unify all this is still a target date, not a shipped artifact. Worth tracking whether it lands in August as hoped, because until it does, "de facto national standard" is still three states and a lot of intent.