Congress Proposes a 3-Year Freeze on State AI Regulation

What the House discussion draft actually proposes

Representatives Jay Obernolte (R-CA) and Lori Trahan (D-MA) released a 269-page discussion draft on June 4 titled the Great American Artificial Intelligence Act of 2026. The bill organizes federal AI policy around four pillars: frontier model governance, workforce monitoring, cybersecurity posture, and AI research and development.

Obernolte described the bill as a direct response to the pace of AI development:

Trahan framed the bill as a security and workforce matter as much as a technology one:

For frontier AI developers, the proposed framework is substantive. Companies with annual revenues above $500 million would be required to publish safety frameworks, report critical incidents within a defined window, and submit to third-party audits every six months. Oversight would be handled by the Center for AI Standards and Innovation, a Commerce Department office the bill would formally establish and fund with $300 million over three years.

The bill is a discussion draft, not introduced legislation. The authors describe it as an invitation for public comment from industry, researchers, and civil society before a formal bill is introduced. That framing matters: the provisions are intended to signal a direction, not to finalize one.

The 3-year preemption clause and what it actually blocks

The safety framework is the visible half of the bill. The less-discussed half is Section 4's preemption clause, which would override state laws "specifically regulating the development" of an AI model for three years from the bill's enactment.

The clause does not preempt laws governing AI use or deployment, only development. But that boundary is contested terrain. California's SB 1047 and its successors, which require large AI developers to conduct safety evaluations before training runs, would fall squarely within "development" regulation. So would dozens of transparency and testing requirements passed by Colorado, Texas, Illinois, and other states since 2024.

As of this writing, more than 30 states have enacted or advanced some form of AI regulation. Most were written in the absence of federal action. The bill would effectively freeze enforcement of state-level development rules while the federal framework is still being designed and staffed. Consumer advocacy groups moved quickly to oppose the preemption clause:

The $500 million threshold that creates a compliance asymmetry

The safety requirements in the bill apply only to companies with annual revenues above $500 million. That threshold covers OpenAI, Anthropic, Google, Microsoft, and Meta. It does not cover most AI startups, many of which are building on top of frontier models or training their own purpose-built systems.

The result is a compliance structure that applies primarily to established incumbents. A startup training a 70-billion-parameter model on proprietary data for enterprise deployment would not meet the $500 million threshold, and would therefore not be subject to the bill's safety framework, even if its outputs reached millions of users. Public Knowledge's Senior Policy Counsel identified a further problem with how the preemption clause was drawn:

This is not an obvious design flaw in the bill's safety framework. The drafters may have reasoned that the highest-risk systems come from the best-resourced labs, and that applying compliance costs to early-stage companies would harm competition. But the threshold also means the framework exempts a large portion of the commercial AI market from any federal safety obligation.

What this bill means for states that have already moved

The most immediate consequence of the bill's passage, if enacted as drafted, would be felt not in Washington but in Sacramento, Austin, and Denver. States that passed AI safety laws with enforcement timelines in 2026 and 2027 would see those timelines suspended. Attorneys general who have announced intentions to enforce AI transparency requirements would be preempted.

The three-year sunset is framed as a temporary measure to allow federal standards to develop. But it creates a window in which neither state enforcement nor federal enforcement is fully operational. The federal Center for AI Standards and Innovation would be standing up its operations during a period when the companies it oversees have no parallel state-level accountability.

The bill's bipartisan structure is genuine. Obernolte and Trahan represent both parties and have worked together on technology policy for several sessions. The safety requirements are not token gestures. But the bill asks states to stand down from enforcement in exchange for a federal framework that does not yet exist, administered by an agency that has not yet been funded, operating under rules that have not yet been written. For companies that have spent two years navigating state-level compliance, that trade has obvious appeal. Whether it produces stronger AI governance or simply buys time is a question the discussion period is designed, but not guaranteed, to answer.