Bipartisan House Lawmakers Unveil Sweeping 269-Page Draft for National AI Framework

At exactly 10:00 a.m. on Thursday, a 269-page PDF materialized on the House Science Committee's server, ending months of speculation about how the US federal government plans to govern artificial intelligence. The "National Artificial Intelligence Framework Act of 2026" represents the most aggressive bipartisan attempt yet to corral a technology that has largely regulated itself. The draft text outlines a comprehensive, mechanism-heavy approach to oversight, shifting the conversation from theoretical existential risks to concrete administrative law.[1]

Instead of banning specific algorithms or use cases, the draft relies on a mechanism of compute thresholds and mandatory disclosures. It establishes a hard line at models trained using more than 10^26 floating-point operations (FLOPs). Any system crossing this computational line would be legally classified as a "Tier 1 Frontier Model." This classification triggers a mandatory 90-day pre-deployment review by a newly proposed Office of AI Innovation, which would be housed within the Department of Commerce.[1]

The framework attempts to solve a growing mechanical problem for the tech industry: the fragmentation of American tech policy. With California, Illinois, and Colorado already advancing their own disparate AI safety mandates, the House draft includes a sweeping federal preemption clause. If passed, it would override state-level risk frameworks, replacing them with a single national standard that dictates how AI companies must operate.[2]

The evidence supporting the need for preemption comes heavily from industry lobbying data. Tech companies argue that complying with 50 different state laws for a globally distributed digital product is technically infeasible and would fracture the US market. The draft explicitly names "interstate digital commerce" as the justification for overriding state laws, a legal mechanism that has historically survived Supreme Court challenges when applied to telecommunications and aviation.[2][4]

The 10^26 FLOP threshold is the bill's most heavily scrutinized metric. According to analysis by the Center for AI Safety, this number captures only the absolute largest models currently in development, such as the next generation of systems from OpenAI, Google, and Anthropic. The vast majority of commercial AI tools, including specialized medical and financial models, would fall well below this line and escape the strictest federal oversight.

However, the scientific consensus on whether compute is an accurate proxy for risk remains highly unsettled. Researchers note that algorithmic efficiency is improving rapidly; a model trained on 10^25 FLOPs in 2026 might possess the same capabilities as a 10^26 FLOP model from 2024. The draft attempts to address this uncertainty by granting the Secretary of Commerce the power to revise the threshold annually, but critics argue this leaves the definition of "dangerous" dangerously fluid.[3]

For the open-source community, the bill introduces a novel legal mechanism: "conditional safe harbor." Developers who release model weights publicly are shielded from downstream liability—such as if a bad actor uses their open-source model to generate malware—provided their models fall below the compute threshold and pass an automated suite of safety benchmarks before upload.

The Electronic Frontier Foundation argues this evidence-pack of benchmarks is fundamentally flawed. Their analysis suggests that automated safety evaluations are easily gamable and disproportionately burden independent researchers who lack the capital to run extensive compliance testing. They claim the 269-page draft essentially creates a regulatory moat for incumbent tech giants who can afford the legal and computational overhead of federal compliance.

Another major claim in the draft involves training data transparency. Section 402 mandates that developers of Tier 1 models maintain a "cryptographically secure ledger" of all copyrighted material used during the training process. This is designed to give authors, artists, and publishers a mechanism to verify if their intellectual property was ingested by a frontier model.[4]

While copyright holders have praised this inclusion, legal experts point out a critical gap in the evidence: the bill does not specify how this ledger interacts with existing fair use doctrine. It mandates recording the data but explicitly avoids stating whether using that data requires compensation, effectively punting the core economic dispute of the AI era back to the federal courts to decide on a case-by-case basis.[4]

To enforce these rules, the bill authorizes $450 million annually to staff the new Commerce Department office. This office would have subpoena power and the authority to levy fines of up to 5% of a company's global revenue for deploying a Tier 1 model without clearance, or for falsifying safety evaluation data during the 90-day review period.[1]

Yet, the practical reality of enforcing these rules remains a massive unknown. Auditing a frontier AI model requires highly specialized talent—talent that currently commands seven-figure salaries in the private sector. Government watchdogs question whether a federal agency, even with a $450 million budget, can attract the necessary engineering expertise to effectively red-team models that are smarter than the tests designed to evaluate them.[3]

The draft also includes provisions for international alignment, requiring the US to establish joint testing protocols with the UK AI Safety Institute and the EU AI Office. This reflects a growing consensus that unilateral AI regulation is porous, as compute clusters can easily be established in non-compliant jurisdictions if domestic laws become too restrictive.[1][3]

The bill now faces a grueling markup process in the House Science Committee. With the 2026 midterm elections looming, bipartisan sponsors are pushing for a floor vote by October, hoping to capitalize on public anxiety regarding AI deepfakes and job displacement before the legislative window closes.[2]

Ultimately, the 269-page draft represents a shift from theoretical debates about AI existential risk to the messy reality of administrative law. It codifies the assumption that AI is too powerful to remain unregulated, but reveals deep uncertainties about exactly which dials the government should turn to control it without stifling the broader technology sector.[1][3]