Great American AI Act: the 269-page bill that could freeze AI regulation in the United States for 3 years
🔎 A moratorium disguised as a framework law
On June 4, 2026, two representatives unaccustomed to working together — Jay Obernolte (Republican, California) and Lori Trahan (Democrat, Massachusetts) — released a 269-page text that could redraw the regulatory map of AI in the United States. The Great American AI Act looks like nothing Washington has produced on the subject so far.
Two days earlier, on June 2, Donald Trump signed an executive order framing voluntary reviews of AI models. The timing is not coincidental: the bill arrives as a legislative complement to this voluntary approach, but with much sharper teeth.
The core of the text? A three-year freeze on all state regulation affecting the development of AI models. During this period, only the federal government would have jurisdiction. California, Colorado, Florida, and a dozen other states that have already adopted or prepared their own laws would be stripped of their regulatory power.
It is an institutional power grab. And it operates through a precise legal mechanism: federal preemption.
The essentials
- The Great American AI Act imposes a 3-year moratorium on state laws concerning the development of AI models, not their deployment or use.
- It creates a Federal AI Evaluation Center responsible for overseeing mandatory audits of frontier models.
- Frontier model developers (OpenAI, Anthropic, Google DeepMind, xAI) would be subject to mandatory safety audits before deployment.
- The text criminalizes impersonation of officials via AI with enhanced penalties.
- The bill follows up on the Trump executive order of June 2, 2026, but goes further by proposing a lasting legislative framework.
- States are massively opposed to it, fearing overly lax regulation at the federal level.
Recommended tools
| Hostinger | Hosting to keep up with AI news | Starting at 2.99 € (June 2026, check on hostinger.com) | Tech content creators |
|---|---|---|---|
The central mechanism: federal preemption for 3 years
Preemption is the constitutional principle whereby a federal law can override a state law on the same subject. The Great American AI Act activates it in a targeted manner: for 36 months from its entry into force, no state could legislate on the development of artificial intelligence models.
The distinction is crucial. The text does not block state laws on the use of AI — for example, regulations on AI in recruiting, healthcare, or criminal justice would remain possible. What is frozen is the regulation of the creation process: training, architecture, training data, and computational resources mobilized.
According to Reuters, this clarification was negotiated foot by foot between the two political camps. Republicans wanted a total freeze; Democrats obtained that deployment remain within the states' jurisdiction.
The problem? This development/deployment distinction is blurry in practice. A model like GPT-5.5 from OpenAI, which dominates the agentic leaderboard with 98.2 points, is continuously updated. The boundary between "developing" and "deploying a new version" becomes legally porous.
Which frontier models are targeted?
The bill explicitly targets "frontier AI models" — models whose capabilities exceed a threshold yet to be defined by the Federal Evaluation Center. In practice, the list of affected companies is short and well identified.
The current models that would fall under this regime are those that reach or exceed the highest thresholds of existing benchmarks. Google's Gemini 3.1 Pro (92 points overall), Anthropic's Claude Opus 4.7 (Adaptive) (90 points, 94.3 in agentic), and xAI's Grok 4.1 (90 points) — would all be subject to mandatory audits.
The text provides that the Federal Center will determine the precise thresholds in consultation with the NIST and the scientific community. But the criteria discussed include: the computing power used in training (measured in FLOPs), performance on specific safety benchmarks, and the ability to act autonomously in digital environments.
It is this last point that is concerning. Agentic models like GPT-5.5 are specifically designed to act autonomously. An AI agent that works while you sleep — planning, executing, iterating — raises fundamental security questions that the bill attempts to regulate.
The Federal AI Evaluation Center: a new bureaucracy?
The Great American AI Act creates an unprecedented entity: a federal center dedicated to the evaluation of AI models, attached to the Department of Commerce. According to FedScoop, this center would have three main missions.
First, to define and maintain the classification thresholds for models (frontier or not). Second, to supervise and certify security audits conducted by independent third parties. Third, to publish quarterly reports on the state of the capabilities of models deployed in the United States.
The originality of the system: audits would not be conducted by the center itself, but by accredited laboratories, modeled on financial certification (SOC 2) or information systems security (ISO 27001). The center would play the role of a supervisory body.
The proposed funding amounts to 450 million dollars over three years, with an annual budget of 150 million starting from the second fiscal year. An amount that critics consider insufficient given the speed of the industry's development. For comparison, the NIST budget for AI standards in 2025 was 34 million dollars.
States Fight Back: California, Colorado, Florida on the Front Lines
The states' reaction was immediate and fierce. On June 5, just 24 hours after the draft's publication, California Attorney General Rob Bonta issued a public statement calling the proposal a "regulatory coup."
California is no minor player in this debate. In 2025, it passed SB 1047, an ambitious law on large model safety, before its governor significantly watered it down. Colorado adopted SB 205, creating a governance framework for high-risk AI systems in sensitive sectors. Florida, for its part, paved the judicial way: as shown by Florida's lawsuit against OpenAI and Sam Altman, states are not waiting for the federal government to act.
The states' argument is twofold. First, legal: the US Constitution protects the police powers of the states, and a three-year moratorium on a subject evolving as fast as AI amounts to a regulatory vacuum, not a legitimate preemption. Second, practical: states believe the federal government lacks the resources to back up its policy, and that the Evaluation Center will be a paper tiger.
The tension is real. The geopolitical context of the chip war between the United States and China adds a layer of complexity. Beijing is blocking H200 deliveries authorized by Washington, and some states fear that overly light federal regulation could compromise long-term national security.
Mandatory audits: what this concretely changes for OpenAI, Anthropic, Google, and xAI
Until now, security audits of AI models were entirely voluntary. OpenAI publishes "system cards", Anthropic publishes safety reports, Google publishes responsibility notes. But none of these documents have any legal weight.
The Great American AI Act would change the game. Developers of frontier models would have to submit their models to an independent audit before any significant deployment. The audit would cover four areas: cybersecurity (can the model's weights be extracted?), dangerous capabilities (can it design biological weapons?), autonomy (can it operate without prolonged human supervision?), and robustness (does it resist attempts to bypass guardrails?).
For companies, the cost would be far from negligible. A full audit of a model the size of GPT-5.5 is estimated between 5 and 15 million dollars by a cybersecurity consulting firm interviewed by Cybernews. This amount remains marginal compared to the training cost (estimated between 500 million and 1 billion dollars for current frontier models), but it introduces an additional delay of 4 to 8 weeks into deployment cycles.
The bill also provides for sanctions: a fine of up to 5% of global revenue for unaudited deployment, and the possibility for the federal government to order a model's withdrawal from the market.
Impersonating government officials: the penalty clause of the text
Amid the 269 pages, one provision has received less attention but deserves examination. The bill creates a specific federal offense for the use of AI to impersonate a federal government official.
The proposed penalty is up to 5 years in prison and a $250,000 fine. This is significantly harsher than existing identity theft laws, which remain applicable but are not suited for real-time generated voice and video deepfakes.
The trigger: a series of incidents that occurred between late 2025 and early 2026, where deepfakes of senior Treasury and SEC officials were used to manipulate financial markets. The phenomenon accelerated with the availability of models like Claude Sonnet 4.6 or DeepSeek V4 Pro capable of generating highly convincing multimodal content at a lower cost.
The text also targets the impersonation of election candidates, but this part remains more vague, as the First Amendment complicates any regulation of political speech, even when synthetic.
Great American AI Act vs EU AI Act : two radically opposed philosophies
Comparison with Europe is inevitable. The EU AI Act, which has been gradually entering into force since August 2025, is based on a classification by risk levels: unacceptable, high, limited, minimal. Each level triggers proportionate obligations.
The Great American AI Act takes a completely different route. Instead of classifying systems by risk, it targets developers by model size. A small model used in a high-risk system (medical diagnosis, for example) would escape federal audits if it does not exceed the frontier threshold. Conversely, a frontier model used to generate poems would be subject to the strictest audits.
| Criterion | EU AI Act | Great American AI Act (draft) |
|---|---|---|
| Approach | Classification by risk of use | Classification by model size |
| Geographic scope | 27 Member States | Federal + preemption of the 50 States |
| Audits | For high-risk systems | For frontier models only |
| Sanctions | Up to 7 % of global revenue | Up to 5 % of global revenue |
| Open source | Partially exempt | No explicit exemption |
| Entry into force | Phased 2025-2027 | 3 years after adoption (estimated late 2026) |
The American approach has a clear advantage: it is simpler to apply. Checking the size of a model is objectively easier than assessing the risk level of each specific deployment. But it has a major flaw: it ignores the systemic risks posed by smaller models deployed at scale.
The Trump context: the executive order of June 2, 2026
The Great American AI Act does not come out of nowhere. It fits into a specific political sequence. The executive order signed by Trump on June 2, 2026, had established a framework of voluntary reviews for AI models, asking companies to submit their systems to a federal evaluation before deployment — without any obligation.
The order was praised by the industry but criticized by Democrats and civil rights organizations as a "toothless tiger". The Obernolte-Trahan bill precisely fills this void: it transforms the voluntary approach into a mandatory one for frontier models, while maintaining the permissive approach for the rest of the ecosystem.
It is a shrewd political compromise. Republicans get a freeze on state laws (which they claim hinder innovation), and Democrats get a federal control mechanism over the most powerful models. The price to pay: a three-year regulatory vacuum over everything that is not "frontier".
Open source in the crosshairs
One point that is making the open source community grind its teeth: the bill does not provide any explicit exemption for open weights models. Initiatives like NVIDIA Nemotron 3 Ultra 550B, presented as the most powerful open-source model in the United States at Computex, could be affected if their parameters exceed the thresholds defined by the Evaluation Center.
The issue is technically complex. An open source model published with its weights can be fine-tuned, modified, combined with other models. How do you audit a system whose end use is unpredictable? The draft proposes that the audit focus on the base model as published, but critics point out that this approach does not cover the risks associated with derivatives.
Open source advocates, such as the Linux Foundation and the Open Source Initiative, have already made it clear that they would oppose any regulation that does not distinguish between the development and distribution of open models.
What the text does not say (and why it's problematic)
A bill is also read in its silences. The Great American AI Act leaves several major gray areas.
First, the exact definition of the frontier thresholds is deferred to a subsequent regulation by the Evaluation Center. This is a potentially dangerous blank check: if the center is underfunded or under political influence, the thresholds could be set too high or too low.
Next, the text says nothing about civil liability. If an audited and certified model causes harm, who is responsible? The developer? The auditor? The deployer? The bill remains silent, leaving common tort law to do the work — or not.
Finally, the issue of training data is sidestepped. The bill does not address either the copyright or the privacy of the data used to train the models. This is a deliberate political choice: these issues are the subject of other ongoing legislation (the NO FAKES Act, copyright reforms). But it means that a model could be "audited" and certified even if its training data were illegal.
❌ Common mistakes
Mistake 1 : Confusing a freeze on development with a freeze on use
Many commentators have presented the text as prohibiting States from regulating AI. This is inaccurate. Only the regulation of development is frozen. A State can still prohibit the use of an AI for credit decisions, for example. The distinction is subtle but fundamental.
Mistake 2 : Thinking the text is already law
This is a draft. A bill published for consultation. It must go through the House Energy and Commerce Committee, then a House vote, then the Senate, then presidential signature. The process will take a minimum of 6 to 12 months, and the text will likely be amended significantly.
Mistake 3 : Directly comparing with the EU AI Act without nuance
The two texts start from philosophically opposed premises. Europe regulates uses; the United States regulates capabilities. Comparing their merits without highlighting this fundamental difference produces a superficial analysis.
❓ Frequently asked questions
Who are the two authors of the bill?
Jay Obernolte, Republican representative from California (former video game developer) and Lori Trahan, Democratic representative from Massachusetts. Their bipartisanship is the most important political signal of the text.
Are open source models like DeepSeek V4 Pro affected?
Yes, potentially. The draft does not provide any open source exemption. If the frontier threshold is set in terms of parameters or FLOPs, an open-weight model like DeepSeek V4 Pro (88 points, Max version) could be subject to it.
What happens after the 3-year preemption?
The bill provides that after 36 months, the Evaluation Center will submit a report to Congress with recommendations for a permanent framework. The freeze could be extended, lifted, or replaced by a permanent certification system.
Can Florida continue its lawsuit against OpenAI?
Yes. The freeze only applies to legislation, not to ongoing lawsuits. Florida's action against OpenAI and Sam Altman remains fully valid.
How does this text affect European companies operating in the United States?
Any company deploying a frontier model on US soil would be subject to audits, regardless of its nationality. A French company like Mistral would have to comply with the same regime as OpenAI if it deploys a model above the threshold in the US.
✅ Conclusion
The Great American AI Act is the most ambitious AI bill ever proposed in the US Congress — and also the most controversial. By trading a three-year freeze on state laws for a federal audit mechanism, it attempts the impossible: satisfying an industry that wants clarity, the states that want to protect their citizens, and the national security agencies that want to stay ahead of China. It will probably achieve none of the above — but the debate it opens will define US AI regulation for the decade to come.