On June 4, 2026, a bipartisan group of House lawmakers released a 269-page discussion draft establishing the first comprehensive federal framework for governing artificial intelligence in the United States. Two days earlier, President Trump had signed an executive order establishing voluntary federal agency reviews of frontier AI models — the same order whose signing ceremony had been cancelled two weeks prior. The legislation and the executive order are designed as complementary instruments. Their combined reception illustrates the political geometry of AI governance in 2026: a broad consensus that something must be done, a persistent disagreement about what, and a specific controversy about who has the authority to do it.
By Vladimir Tsakanyan, PhD · Center for Cyber Diplomacy and International Security · cybercenter.space
The Great American Artificial Intelligence Act of 2026 is, in terms of its scope and structural ambition, the most significant AI governance legislation introduced in the United States Congress. Its 269 pages address frontier model oversight, cybersecurity, open-source software security, workforce impacts, AI literacy, research funding, criminal penalties, whistleblower protections, international standards, and the governance of AI in federal procurement — a legislative architecture whose breadth reflects the recognition, shared across its bipartisan sponsorship, that artificial intelligence has become a system-level challenge requiring a system-level response.
Its introduction by Representatives Jay Obernolte and Lori Trahan, joined by four co-sponsors, was accompanied by near-universal criticism from labour unions, consumer advocacy organisations, and the House Democratic Task Force on Artificial Intelligence — not for the substance of its AI governance provisions, which drew broad acknowledgment as a serious legislative effort, but for a single structural mechanism whose implications extend well beyond the bill’s stated scope: the three-year preemption of state laws governing the development of AI models.
Understanding the bill requires engaging with both dimensions — the substantive governance architecture it proposes and the political and constitutional controversy that architecture has generated — and recognising that the two are more closely connected than the initial reception suggests.
The Four Pillars and Their Governance Logic
The bill organises its provisions around four pillars, each addressing a distinct dimension of the AI governance challenge and each reflecting a specific legislative judgment about where federal intervention is most necessary.
Frontier model governance is the bill’s centrepiece. Large frontier AI developers — defined as those with more than five hundred million dollars in annual revenue — would be required to develop and publicly disclose risk frameworks assessing the potential for their models to contribute to critical harm, defined in the bill as the death or serious injury of at least one hundred people or economic damage of at least one billion dollars. The frameworks must address four specific risk categories: cybersecurity vulnerabilities, biosecurity risks, chemical biological radiological and nuclear uplift potential, and loss-of-control scenarios — the four categories that the AI safety community has identified as the most consequential near-term risks from frontier models.
Before releasing a new frontier model, covered developers must implement their risk frameworks and disclose any material deviations. Critical safety incidents — events suggesting that a deployed model poses risks not anticipated in the development framework — must be reported to the Center for AI Standards and Innovation. Developers must maintain whistleblower protections for employees who raise concerns about safety, security, or compliance issues.
The enforcement mechanism that distinguishes this framework from purely voluntary approaches is the mandatory independent verification organisation regime. Covered developers must retain IVOs — private-sector entities certified by CAISI and explicitly prohibited from having financial or organisational ties to the companies they audit — to conduct semi-annual assessments of their compliance with the framework requirements. IVO audit reports must be submitted to CAISI and are available to state attorneys general upon request.
Cybersecurity provisions address both the AI threat to cybersecurity and the cybersecurity vulnerability of AI systems themselves. The bill authorises CISA to award grants to maintainers of widely used open-source software for security improvements including patching, maintenance, and security audits — a specific legislative response to the documented vulnerability of open-source software that represents the majority of the code base on which modern digital infrastructure runs. The GAO would evaluate security protocols protecting AI model weights and the security of the open-source ecosystem more broadly. The Cybersecurity Information Sharing Act of 2015 would be extended through 2035, providing continuity for the information sharing architecture that the legislation’s cybersecurity provisions depend on. The bill also establishes federal criminal penalties for AI-assisted financial crimes and for the use of AI to impersonate government officials.
Workforce provisions require large frontier AI developers to provide WARN Act disclosures for AI-related workforce reductions — the first time federal legislation would formally link AI adoption to the labour displacement notification requirements that govern mass layoffs. The Census Bureau and Bureau of Labor Statistics would be directed to revise federal surveys to incorporate AI use and adoption questions, creating the data infrastructure required to measure AI’s actual impact on the labour market. A Labour Department AI Workforce Research Hub would be established, and education provisions would fund K-12 AI literacy programmes, NSF scholarships for AI-related fields, and grants to broaden participation in AI research.
Research and development provisions formally codify the National Artificial Intelligence Research Resource — a Biden-era programme providing researchers with access to computing infrastructure and data — and authorise NIST and the National Science Foundation to establish grants and prizes for AI research, education, and workforce development. CAISI would be formally codified as the federal government’s primary AI standards institution, with an authorisation of one hundred million dollars per year from 2027 to 2029, and would lead international AI standards development with an explicit mandate to exclude Chinese government participation.
Analyst note
The explicit exclusion of Chinese government participation from NIST-led international AI standards development is among the bill’s most geopolitically significant provisions and the one that has received the least analytical attention in the initial coverage. International AI standards — for model evaluation, for risk assessment, for safety testing — will become the technical foundation on which AI governance frameworks around the world are built. A standards development process from which China is explicitly excluded produces standards that reflect the governance values and technical priorities of the excluding coalition. Standards from which China is excluded are also standards that China has no incentive to adopt, creating a bifurcated international AI standards environment that amplifies the broader digital governance fragmentation documented in the WEF’s 2026 assessment. The provision is a coherent competitive strategy. Its implications for the international governance architecture of AI are more complex than its framing as an innovation protection measure suggests.
The IVO Mechanism: The Bill’s Most Technically Significant Innovation
The independent verification organisation regime is the most technically novel element of the bill and, if implemented effectively, the most consequential departure from the voluntary self-assessment model that has characterised US AI governance to date.
The IVO framework creates a new category of regulated professional entity — private-sector organisations certified by CAISI to conduct semi-annual audits of frontier AI developers’ compliance with their risk frameworks. The structural requirements for IVO certification include independence from the companies being audited, technical competence in the relevant risk categories, and accountability to CAISI for the quality and accuracy of their assessments. The audit reports produced by IVOs are not merely internal compliance documents. They are regulatory submissions to CAISI and are available to state attorneys general on request — a disclosure mechanism that creates a form of public accountability without requiring the full public disclosure of proprietary model information.
The bill’s definition of what IVOs must assess is specific and challenging: compliance with the developer’s framework across the four risk categories, and an independent determination of whether the framework achieves acceptable levels of risk mitigation. This second requirement — the adequacy determination — is more demanding than a pure compliance check. It requires IVOs to form an independent technical judgment about whether the developer’s approach to risk is sufficient, not merely whether the developer has followed its own stated approach.
The governance of the IVO ecosystem raises questions that the discussion draft has not fully resolved. Who develops the technical standards against which IVO assessments are conducted? CAISI is tasked with certifying IVOs but the bill’s provisions on the standards CAISI would develop for this purpose are less specific than the requirements for the IVOs themselves. The adequacy determination requires a baseline — a definition of what constitutes acceptable risk mitigation in each of the four categories — whose development is a technical and policy challenge of the first order, requiring expertise in AI safety, cybersecurity, biosecurity, and catastrophic risk that no existing institution has fully assembled.
The conflict of interest provisions — requiring that IVOs have no financial or organisational ties to the companies they audit — address the most obvious governance risk in a third-party audit regime. They do not address the more subtle dynamics of a small professional ecosystem in which the pool of technically qualified auditors is concentrated in a community with prior professional relationships with the companies being assessed. The accounting profession’s experience with the management of auditor independence provides relevant precedent — precedent whose lessons include the inadequacy of formal independence rules in the absence of structural mechanisms that limit the practical concentration of audit relationships.
The Preemption Controversy and Its Structural Logic
The three-year preemption of state laws governing the development of AI models is the provision that has generated the strongest opposition and the provision whose structural logic is most difficult to assess without reference to the alternative it is designed to address.
The opposition to preemption is grounded in a concrete concern: more than forty states have introduced AI legislation in 2025 and 2026, and several have enacted laws that the preemption would prevent from taking effect. California’s AI worker protection framework, Colorado’s Consumer Protections for Artificial Intelligence Act — effective June 30, 2026 — and New York’s AI transparency requirements are among the state laws that would be affected. The characterisation by one advocacy organisation that the preemption would be “a generational mistake for consumer protection” reflects the view that state laws represent genuine and necessary protections that a federal framework has not yet provided and may not adequately substitute for.
The preemption’s defenders, including industry bodies and the bill’s sponsors, make a structural argument: a regulatory environment in which AI model developers face fifty different and potentially conflicting state requirements for how they build their models creates a compliance architecture that primarily advantages large incumbents — who can manage compliance complexity — over the smaller competitors and open-source developers that the federal framework explicitly exempts from its most onerous requirements. Representative Erin Houchin’s statement that a patchwork of fifty different state laws would make it harder for American companies to innovate while doing little to improve consumer protections reflects this structural concern.
The bill’s preemption is more carefully scoped than the initial criticism suggests. States retain full authority to regulate the use and deployment of AI systems within their borders — including laws covering civil rights, labour and workplace protections, copyright, the production of child sexual abuse material, and consumer privacy. The preemption applies specifically to the development of AI models: how they are built, tested, and evaluated before release. States cannot, under the bill’s preemption, impose development-phase requirements that differ from the federal framework. They can impose any requirements they choose on how AI is deployed and used within their jurisdictions.
The significance of the use-development distinction depends on whether the most consequential AI harms arise from decisions made during development — the architecture of the model, the training data, the safety evaluation — or from decisions made during deployment — the contexts in which the model is applied, the safeguards that deployers put in place, the populations to whom it is made available. Reasonable assessments of where the most consequential governance decisions occur support both positions, and the bill’s implicit judgment that development-phase federal governance is the priority is a policy choice that the legislative record does not fully justify.
Analyst note
The preemption provision’s relationship to the California regulatory agenda deserves specific attention given California’s status as the most consequential state AI regulator and the home jurisdiction of the largest concentration of frontier AI development. Governor Newsom’s June 2026 executive order on AI workforce impacts — directed at the deployment and workforce consequence dimensions of AI adoption — would survive the preemption because it addresses workforce impact rather than model development. The California RAISE Act and equivalent development-phase bills — which the preemption is most directly designed to address — would be frozen. The practical effect is to preserve California’s most ambitious deployment-phase governance agenda while temporarily preventing its development-phase counterpart, creating a regulatory environment in which the same frontier AI developers face both a federal development-phase framework and California’s deployment-phase requirements simultaneously.
The Relationship to the Executive Order
The bill’s introduction two days after President Trump signed the executive order on AI innovation and security is not coincidental. The two instruments are designed to operate as complementary components of a layered federal AI governance architecture — the executive order establishing voluntary mechanisms and directing agency action, the legislation providing the statutory authority and mandatory requirements that the executive order cannot supply.
The executive order of June 2, 2026 established a voluntary AI cybersecurity clearinghouse coordinating vulnerability scanning, validation, and patch remediation across the AI industry and critical infrastructure operators. The bill’s mandatory IVO audit regime, its critical safety incident reporting requirements, and its CISA open-source security grants all build on and extend the executive order’s voluntary framework into binding legal requirements. The executive order directed CAISI to develop standards; the bill codifies CAISI and funds its work. The executive order established voluntary pre-release review windows for frontier models; the bill establishes mandatory risk frameworks and semi-annual audits.
The design logic of this layering is straightforward: the executive order moves immediately, within the scope of executive authority, to establish the institutional infrastructure and voluntary practices that the governance architecture requires. The legislation provides the statutory foundation that makes the architecture durable, the mandatory requirements that give it enforcement teeth, and the appropriations that provide its institutional funding beyond what the executive branch can allocate without congressional authorisation.
The practical question is whether the legislative process will produce a bill that is sufficiently similar to the discussion draft to preserve this design logic, or whether the controversy over preemption and the political dynamics of the 118th Congress will produce amendments that alter the architecture in ways that undermine its coherence. The sponsors gave no timeline for formal introduction of the bill, and the discussion draft designation signals that the text is intended to generate reactions that will inform the formal bill rather than to represent a finalised legislative position.
The Open-Source Dimension
The authorisation of CISA grants for open-source software security — a provision that has received less attention than the frontier model and preemption provisions — addresses a vulnerability that the 2026 incident record has documented with unusual specificity.
Open-source software represents the majority of the code base on which modern digital infrastructure, including AI systems, depends. Its security has historically been maintained by volunteer communities and a small number of commercial entities whose investment in security has been inconsistent with the critical role the software plays in global infrastructure. The npm package compromise of March 2026, the sustained exploitation of known vulnerabilities in widely deployed open-source libraries, and the upstream supply chain attacks documented in multiple 2026 breach reports all reflect the same structural condition: widely used open-source software is maintained without the security investment that its criticality warrants.
The CISA grant authorisation addresses this condition directly. By providing federal funding for security improvements — patching, maintenance, and security audits — to maintainers of widely used open-source software, the bill creates a mechanism through which the externality problem of open-source security can be partially addressed through public investment. The maintainers who bear the cost of security work receive compensation that the market has not provided. The infrastructure whose security is critical to the entire digital economy receives the investment that its role requires.
The provision’s effectiveness depends on implementation details that the discussion draft does not fully specify: the criteria for which open-source projects qualify as “widely used,” the process through which grants are evaluated and awarded, the security standards that grant-funded work must meet, and the mechanism for verifying that security improvements have been implemented effectively. CISA’s capacity to administer a grant programme of this kind — at the current state of its institutional capacity following the workforce reductions of 2025 — is a relevant consideration that the bill’s sponsors have not addressed publicly.
The International Governance Dimension
The bill’s international implications extend beyond the explicit China exclusion provision to its potential effects on the emerging global AI governance architecture.
The EU AI Act, in full effect from August 2026, establishes a risk-based framework for AI governance that has influenced AI governance thinking across multiple jurisdictions. The Great American Artificial Intelligence Act’s approach differs from the EU framework in its focus on frontier model development rather than deployment-phase risk categories, in its use of private-sector IVOs rather than public regulatory bodies for compliance verification, and in its explicit framing around national security and competitive advantage rather than fundamental rights.
The coexistence of these two major AI governance frameworks — one US-based, one European — will define the international regulatory environment for frontier AI developers that operate in both markets. Developers subject to both frameworks will face potentially overlapping and potentially conflicting requirements for risk assessment, audit, and incident reporting. The mechanisms for regulatory cooperation between CAISI and its EU counterparts — the AI Office established under the EU AI Act — are not addressed in the bill, and the establishment of mutual recognition arrangements or coordinated audit frameworks would require diplomatic engagement that the bill does not direct.
The explicit mandate to exclude Chinese government participation from NIST-led international AI standards development creates a further dimension of international governance complexity. Standards bodies including ISO, the ITU, and the IEEE have historically operated on principles of open participation that the bill’s mandate challenges. The practical effect of the exclusion — whether it results in bifurcated international standards or in Chinese non-participation in US-led standards that the rest of the world adopts — will depend on diplomatic and institutional dynamics that extend well beyond the bill’s provisions.
Bottom Line Assessment
The Great American Artificial Intelligence Act of 2026 represents the most substantive attempt yet by the US Congress to establish a durable federal framework for AI governance. Its IVO audit regime, its risk framework requirements, its open-source security grants, and its cybersecurity provisions address real and documented governance gaps with mechanisms that reflect serious legislative effort.
Its reception — near-universal criticism of the preemption provision from consumer advocates, labour organisations, and some Democratic members — reflects the intensity of the political contest over who controls AI governance in the United States, and the specific concern that a three-year freeze of state development-phase regulation will benefit large developers at the expense of the consumer and worker protections that state legislatures have been attempting to provide.
The bill is a discussion draft without a formal introduction timeline. Its provisions will evolve through a legislative process whose outcome is uncertain and whose timeline, given the complexity of the political dynamics it has already generated, is not predictable. What the discussion draft establishes, regardless of its legislative fate, is a reference point for the federal AI governance debate — a detailed architecture of what comprehensive federal AI regulation could look like — that will shape the terms of that debate through the remainder of the legislative session and beyond.
The most consequential questions the discussion draft raises are not about its specific provisions but about the governance infrastructure it requires. An IVO regime requires auditors with expertise in frontier AI safety, cybersecurity, biosecurity, and catastrophic risk. CAISI requires institutional capacity to certify those auditors, develop the adequacy standards against which their assessments are measured, and process the audit reports it will receive. An open-source security grant programme requires CISA to administer a new function at a moment when its institutional capacity is operating below its established baseline.
The gap between the governance architecture the bill proposes and the institutional capacity available to implement it is the central challenge that its sponsors will need to address as the discussion draft moves toward a formal bill. The architecture is serious. Its realisation depends on institutional investments that the bill’s authorisations alone cannot guarantee.
Great American Artificial Intelligence Act · AI Governance · Frontier Models · IVO Audit Regime · State Preemption · Open-Source Security · CAISI · CISA · Trump AI Executive Order · International AI Standards · Vladimir Tsakanyan
Leave a comment