OpenAI Pushes ‘Reverse Federalism’ to Build US AI Safety Rules

OpenAI published a detailed policy argument on July 15 laying out how it believes the United States should govern the most powerful AI systems — and its strategy hinges on states doing the heavy lifting until Congress catches up.

Written by Chris Lehane, OpenAI’s chief global affairs officer, the piece introduces a framework the company calls “reverse federalism”: the idea that if enough states pass substantially similar AI safety laws, they effectively create a national compliance standard that Congress can then adopt wholesale. California, New York and Illinois are already there, OpenAI argues, and together they represent an estimated 40% of the U.S. AI market — enough weight to function as a de facto national regime even without a single federal law.

What the Three-State Baseline Actually Requires

OpenAI identifies three elements it says must appear in any state law for this strategy to work. First, AI developers must produce documented safety frameworks, including risk assessments for frontier models and public disclosure of results. Second, companies must report serious safety incidents. Third, they must submit to independent, objective audits. OpenAI describes California as having “established the core disclosure framework,” New York as proving “the approach could be adopted across jurisdictions,” and Illinois as “requiring independent verification of key disclosures.”

The company is explicit that additional provisions beyond those three — often added to secure enough legislative votes — risk creating the regulatory patchwork it is trying to avoid.

“Without that discipline, we risk policy creep and a patchwork of state laws that are difficult for regulators to enforce, confusing for consumers to navigate, and divert developer resources—especially at start-ups and small companies—that would be better invested in safety,” wrote Lehane.

The Federal Picture: A Cybersecurity Deadline and a Bipartisan Bill

At the federal level, OpenAI signals it is engaged with the Trump Administration on a framework specifically for government cybersecurity testing of the most capable AI models, with a target completion date of early August 2026. The company also endorses — cautiously — the Great American AI Act, a bipartisan House discussion draft introduced by Reps. Jay Obernolte and Lori Trahan, calling it “a productive step forward” while stopping short of full-throated support for every provision.

OpenAI also calls for strengthening the Center for AI Standards and Innovation (CAISI), a federal body created under President Biden and continued under President Trump, as the durable institutional home for frontier model evaluations. CAISI evaluations already cover cybersecurity, biosecurity and chemical-weapons risks, with some work conducted in classified settings — a scope that requires specialized technical skills and, in some cases, security clearances.

The international dimension is equally ambitious. CEO Sam Altman published a proposal in the Financial Times calling for “a US-led international forum that establishes accepted standards, provides expert and impartial analysis of capabilities and risks, and makes the technology available to nations and companies that participate and follow the rules.”

Lehane’s post frames the domestic regulatory momentum as groundwork for exactly that kind of global framework — positioning U.S. standards as the template other democracies adopt rather than the other way around.

OpenAI Isn’t Alone — and There Is Real Tension Here

The announcement does not land in a vacuum. Google DeepMind CEO Demis Hassabis has similarly called for an independent expert body to evaluate advanced AI before release, though he envisions a new organization rather than building on an existing government agency like CAISI. Both OpenAI and Anthropic supported Illinois SB 315 as it moved through the state legislature, where it passed with broad bipartisan backing.

But the federal bill has run into meaningful opposition. Labor unions, consumer advocacy groups, and a formal House Democratic commission have pushed back hard on the Great American AI Act’s three-year preemption clause, which critics argue would freeze state and local AI protections while leaving oversight to a federal government that has not yet demonstrated the capacity to enforce them. OpenAI threads this needle carefully — endorsing federal action while insisting states should remain active partners, not sidelined actors.

Why This Matters If You Are Building a Career Right Now

Every compliance function embedded in these laws — risk assessments, audit regimes, incident reporting pipelines, whistleblower protections — needs people to staff it. AI policy is already one of the fastest-growing career tracks of 2026, and demand is outpacing the supply of qualified candidates at frontier labs, federal agencies, law firms, think tanks and international organizations.

For law students, the enforcement architecture alone points to a sustained wave of legal work: the proposed federal bill includes civil penalties up to $1 million per violation per day, state enforcement by attorneys general, and mandatory audit requirements. For computer science students, CAISI’s classified evaluation work signals that federal technical roles are growing and will increasingly require both AI engineering expertise and security clearances. And because many safety, governance and policy roles hire on demonstrated skills rather than specific credentials, students from a wide range of backgrounds — policy, ethics, public administration, even communications — are competitive candidates.

Students entering the workforce over the next 1-3 years will be among the first trained during this regulatory buildup. The infrastructure being debated right now is the infrastructure they will spend their careers operating inside.

Source: OpenAI