- Lead. Representatives Lori Trahan (D-Mass.) and Jay Obernolte (R-Calif.) released a 269-page discussion draft on June 4 that, for the first time, proposes a comprehensive federal AI governance framework in the United States — binding the largest frontier model developers to safety-incident reporting while preempting conflicting state-level rules on how those models are built.
- Fact. The bill targets “large frontier developers” — companies earning more than $500 million in annual revenue from AI — and would require them to report critical safety incidents to the federal government, creating a disclosure mechanism that does not currently exist in US law.
- Stake. The preemption clause is the most contested provision: states retain authority to regulate how AI systems are used, but lose the ability to legislate how the most powerful models are built — a distinction that would effectively nullify several pending state bills and directly reverses the approach taken by states like Colorado that tried to regulate at the development layer.
The bipartisan discussion draft, titled the Great American Artificial Intelligence Act (GAAIA) and co-sponsored by Representative Erin Houchin (R-Ind.), was released on June 4 as a 269-page working document intended to solicit feedback before formal introduction, FedScoop reported. The bill represents the most substantive attempt yet by Congress to move beyond sector-specific AI patches — such as the deepfake TAKE IT DOWN Act — toward a unified federal framework.
What the bill proposes
The draft is organised into four titles: Frontier AI Governance, Workforce, Cybersecurity, and Research, Development and International Cooperation. The governance title contains the most commercially significant provisions. Large frontier developers — those with annual AI-related revenues exceeding $500 million and the training compute to match — would be required to report “critical safety incidents” to a designated federal authority within a specified timeframe. The bill does not define the precise taxonomy of reportable incidents, which advocates for stronger oversight have flagged as a gap that would need to be filled before the draft advances.
On funding, the bill authorises $100 million per fiscal year for fiscal years 2027 through 2029 for a Centre for AI Standards and Innovation — a relatively modest sum given the scale of the industry it would oversee. A separate provision directs the Government Accountability Office to evaluate federal AI adoption progress and identify existing regulations that unnecessarily burden AI infrastructure, a nod to the administration’s broader deregulatory posture.
Ancillary provisions include penalties for AI systems that impersonate government officials, a requirement that the Census Bureau and Bureau of Labor Statistics add AI adoption questions to their periodic surveys, and authorisation for testbed programs and international standards coordination — the latter responding to European and British regulatory advances that US firms have argued are being shaped without adequate American input.
The state preemption fight
The preemption architecture will determine whether the bill advances. Representative Houchin warned against allowing “fifty different state laws” that would “make it harder for American companies to innovate and compete” — a line that positions the bill as a defence of large frontier labs against state-level fragmentation as much as a governance measure. Representative Trahan emphasised the opposite framing, arguing the framework would address AI threats “without smothering American innovation.”
The preemption debate is not theoretical. Colorado gutted its own AI law weeks before it was due to take effect, partly in anticipation of exactly this federal intervention. The GAAIA would prevent other states from following Colorado’s original approach of regulating at the model-development layer, while allowing post-deployment oversight — consumer protection, bias auditing, discriminatory-outcome liability — to remain at the state level. Whether that division is coherent in practice, given that deployment and development decisions interact at every stage of a model’s lifecycle, is a question the discussion draft does not fully resolve.
What comes next
As a discussion draft, the GAAIA has no scheduled committee vote. Sponsors have invited public comment and stakeholder feedback before any formal introduction, meaning the bill’s timeline through committee markup and floor votes remains undefined. The 119th Congress must complete its work by January 2027, setting a practical deadline. Given the pace of AI legislative activity in the Senate — where several competing frameworks remain in early stages — a bicameral conference process before the end of the year would be unusually fast, making the most likely near-term outcome a series of hearings that sharpen the bill’s contested provisions rather than final passage.