On June 4, 2026, House Energy and Commerce Committee principals Rep. Jay Obernolte (R-CA) and Rep. Lori Trahan (D-MA) released a 269-page discussion draft that puts the Supremacy Clause, the Commerce Clause, and the next decade of American AI policy on a collision course. The "Great American Artificial Intelligence Act of 2026" would, for three years, freeze any state law that "specifically regulates the development" of frontier AI models [1]. Twenty-six days later, on June 30, the Colorado AI Act begins enforcement against the same companies. California, in turn, has signaled it will litigate. The constitutional question is no longer academic. It is calendared.
Pedantic but necessary point: this is a discussion draft, not an introduced bill. The distinction matters in preemption analysis. Discussion drafts carry no Congress.gov bill number, are not eligible for floor consideration, and function as vehicles for stakeholder feedback to GAAIA@mail.house.gov before formal introduction. The four co-sponsors named in the June 4 press materials are Rep. Scott Franklin (R-FL-18), Rep. Suhas Subramanyam (D-VA-10), Rep. Erin Houchin (R-IN-09), and Rep. Scott Peters (D-CA-50) [2][3][6]. The preemption clock in the bill, however, is statutory, and would be enforceable the moment it became law. That is why state attorneys general are already mobilizing.
What the 269 pages actually do
The draft is structured in three operative parts. Part one imposes a three-year moratorium on state laws that "specifically regulate the development" of AI models, paired with an express reservation of state authority over the use and deployment of AI systems in employment, housing, healthcare, and credit [1][4]. The use-versus-development distinction is the bill's central federalism bargain. States can keep regulating what AI does in the lives of their residents, but for three years they cannot regulate how the underlying model is built.
Part two defines and regulates a class of "covered large frontier developers" as those with more than $500 million in annual revenue in the prior calendar year. The threshold is keyed to revenue, not compute, which is a deliberate drafting choice; the sponsors have named Anthropic, OpenAI, xAI, and Google DeepMind as entities the threshold would capture [1]. Covered developers must publish a "Frontier AI Safety Framework" identifying catastrophic-risk thresholds, retain a NIST-licensed "Independent Verification Organization" (IVO) for semi-annual third-party audits covering cybersecurity, biosecurity, CBRN uplift, and loss-of-control scenarios, and submit audit findings to the Center for AI Standards and Innovation (CAISI) inside the Commerce Department [1][5]. The draft treats audit results as confidential, with a redacted summary shared with CAISI [5].
Part three is the institutional core. CAISI would be formally codified and authorized at $100 million per fiscal year for FY2027 through FY2029, roughly seven times its current $15 million annual appropriation [1]. The bill also renews the Cybersecurity Information Sharing Act of 2015 through fiscal 2035, a textual signal that Congress is treating the IVO regime as a national-security instrument rather than a consumer-protection one [1]. CAISI is barred from maintaining formal ties to the companies it monitors, a structural guardrail against the industry-capture risk that civil-society critics have flagged.
Enforcement runs through civil penalties of up to $1 million per day per violation for non-compliance or material misrepresentations to auditors, paired with a 24-hour reporting window for imminent catastrophic risks and a 15-day window for critical safety incidents [1][8]. Whistleblower protections for AI employees who report violations sit alongside the penalty regime. The $1 million per day ceiling and 24-hour reporting window are notable features of the federal enforcement regime.
The constitutional collision, dated
Federalism stakes in this bill are unusually concrete. Colorado's AI Act, SB24-205, begins enforcement on June 30, 2026, which means a frontier developer subject to both the Colorado statute and the federal preemption window is, on day one of the moratorium, in a position where one sovereign says comply and the other says wait [12]. Reporting on the Trahan office FAQ has identified the collision: California's AB 2013, which requires training-data summaries, would be preempted, as would portions of California SB 942 on content watermarking [4]. IAPP has read the bill's "specifically regulate the development" language as narrower than California's drafters intended [7], but narrower is not narrow enough for California, which has been identified as likely to challenge the preemption in court [4].
The bipartisan wrinkle is real, and it is the most underreported feature of the bill. The Obernolte-Trahan draft is a fundamentally different preemption vehicle from the December 11, 2025 Trump executive order titled "Ensuring a National Policy Framework for AI" [14]. The December EO does not preempt by statute. It directs the Department of Justice to bring preemption challenges against state AI laws that "unduly burden" federal AI policy, conditions federal funding on state cooperation, and was issued under Article II authority that constitutional scholars and state AGs have argued is exceeded [11][12][14]. Twenty-two state attorneys general, joined by the District of Columbia, have pushed back on federal preemption efforts, with two of those attorneys general being Republicans [6]. NPR has tracked the bipartisan pushback from governors and AGs of both parties [13].
Statutory preemption under the Supremacy Clause is the doctrinal bet, and the bet is that Congress can do what the President alone arguably cannot: set the terms on which federal law displaces state law. That is more durable as a matter of constitutional structure, but it requires 218 votes in the House and 60 in the Senate. The December EO required none of those votes, and that asymmetry is precisely why the bill is structured as a discussion draft. Trahan, who carries credibility with labor and consumer advocates, is the principal sponsor most plausibly positioned to deliver the Democratic votes the bill would need at introduction, and her credibility is now being strained by her endorsement of preemption [2][9]. Obernolte, a former software engineer and chair of the House AI Task Force, brings technical fluency but a more limited progressive coalition [2][3].
The carve-outs, and why they will be litigated
The bill's preemption is not a clean sweep. The sponsors' FAQ preserves state authority over civil rights enforcement, labor protections, copyright, child sexual abuse material prohibitions, and consumer privacy, and the text carves out three subject-matter areas the bill labels as "children, election integrity, sexual deepfakes" [1][6]. The carve-outs are textually broad, particularly the "election integrity" reference, and operationally ambiguous. Public Citizen and Americans for Responsible Innovation have argued the carve-outs are too narrow because most state AI law targets deployment rather than development, and the "specifically regulate the development" phrase does not reach those statutes [9]. Americans for Responsible Innovation president Brad Carson has called the preemption provision a "generational mistake" and warned that the bill "takes the current floor on state AI legislation and turns it into a federal ceiling" [4][9]. NetChoice director of policy Patrick Hedger has raised a separate concern about the IVO regime, warning that "the bill's aggressive auditing regime and data-sharing requirements for developers could risk entrepreneurs' trade secrets and private records" [4].
A broader coalition of 22 state attorneys general, joined by the District of Columbia, has already pushed back on the federal preemption track, with two of those attorneys general being Republicans [6]. Industry support, by contrast, has been more measured. The Center for Data Innovation has called the bill "one of the most serious federal attempts yet to establish a coherent national AI framework" and has praised the bill's "transparency requirements for the most powerful AI systems, independent auditing through licensed verifiers, and a federal standard that prevents conflicting mandates from fragmenting the national AI ecosystem" [10]. Critics note the inverse: the bill is also a regulatory haven for the four named firms, removing the state-level competition that has, in the absence of federal action, forced disclosure and risk-management requirements onto frontier developers in Colorado and California. Both characterizations can be true at once, and the bipartisan framing of the bill depends on pretending they cannot both be.
The model-suspension precedent
The bill lands in a federal enforcement landscape that is already shifting. On June 12, 2026, the US government ordered Anthropic to suspend Fable 5 and Mythos 5 model access for foreign nationals, via a Commerce Secretary Lutnick letter to CEO Dario Amodei, marking the first time a government has forced a leading AI company to take a publicly deployed frontier model offline [15]. That action, taken under existing Commerce Department authority, sits in an uneasy relationship with the Obernolte-Trahan draft, which would formalize a new audit-and-reporting regime for covered frontier developers without addressing the executive-branch suspension power that the Lutnick letter just exercised. The bill's text does not preempt or codify that authority, and it does not address whether the IVO-audit findings that CAISI would receive could be used to justify future model-suspension orders. New York Governor Hochul's pending decisions on state AI legislation are, separately, "especially consequential for whether a national or state-by-state framework ultimately governs AI in the United States" [15], and the New York legislature has been working on its own foundation-model transparency regime that would sit in the same preemption crosshairs as California's AB 2013 if the federal bill becomes law.
The road to introduction
The Obernolte-Trahan draft is the most serious federal preemption effort on AI to date, and the most vulnerable. The bipartisan cover is real at the principal-sponsor level but conditional below it. Subramanyam and Peters, the two Democratic co-sponsors most identified with state-level AI accountability work, may withdraw support if the carve-outs are narrowed rather than widened during the comment period [2][9]. The Trump December EO remains in litigation, and the statutory path the bill pursues is the more durable path to preemption that can survive the Article II challenge. The clock runs against all of it: Colorado's June 30, 2026 enforcement date, the comment inbox at GAAIA@mail.house.gov, and an election cycle that will not wait on a 269-page bill.
The constitutional question is no longer whether it will be answered. It is which branch of government, and which level of federalism, gets to answer it first. The Trahan-Obernolte wager is that Congress can pre-empt by statute what the executive can only challenge by litigation. State AGs are betting the Supremacy Clause cuts the other way.