A week ago, I wrote about the AI Executive Order the White House dropped on June 2, the one framed as an AI security order that was actually more about the government getting its own house in order than regulating the private sector. Since then, three more documents have landed.
The new additions are NSPM-11, the National Security Presidential Memorandum on AI in the national security enterprise; a House discussion draft called the Great American AI Act, introduced by Rep. Obernolte and co-sponsored on a bipartisan basis; and a separate bill from Rep. Van Duyne called the CREATOR Act, which is specifically about AI-generated visual art. None of these are law yet. NSPM-11 is already in effect as a presidential directive. The two House bills are discussion drafts, meaning they are published to generate feedback before formal committee consideration; they could change significantly or not move at all. Read them with that in mind.
What is NSPM-11?
NSPM-11, signed June 5, is directed entirely at the national security enterprise, DoD, the intelligence community, and agencies with a national security function. It doesn’t touch private companies or civilian government systems except where those systems interact with the national security apparatus. If you work in defense contracting, that last part matters. The memorandum includes a provision directing agency heads to terminate contracts, including subcontracts, with companies that demonstrate a repeated pattern of conduct inconsistent with the memo’s core policies. Those policies, organized under four pillars labeled adoption, adaptation, assurance, and accountability, center on the idea that AI used in national security contexts must be reliable, controllable, and consistent with constitutional protections. Agencies can grant waivers, but those waivers are capped at one year and have to be reported in writing to the White House. That is a meaningful accountability mechanism, at least on paper.
The assurance pillar is the one with the sharpest edge from a security standpoint. It requires, through contract or other means, that no commercial vendor or adversary has the ability to disable, degrade, or materially modify an AI system that military or intelligence personnel are depending on without the government’s knowledge and approval. Think about what that means in practice: the government is essentially demanding a kill-switch veto on any AI it becomes operationally dependent on. That is not a trivial ask from a vendor’s perspective, and it will reshape how AI contracts get structured in the defense and intelligence space over the next few years.
The memo also rescinds NSM-25, the prior administration’s AI national security guidance. That is worth noting because it signals a deliberate clean break from the previous governance framework, not a continuation or refinement of it.
What is The Great American AI Act?
The Obernolte bill is the more consequential document from a regulatory standpoint, and Section 121 is the part I’d tell any compliance or legal team to read first. It federalizes AI model regulation, meaning it would preempt state laws that regulate how frontier AI models are developed. If this passes in anything like its current form, the patchwork of state AI laws that has been building, including laws in states like Texas and California, gets pushed aside for anyone in the frontier model space.
The bill defines a frontier model with some precision: it’s a foundation model trained using more than 10^26 operations. That’s a very large number, and it’s not a threshold that applies to most AI tools in common use. The governance obligations in Title I only apply to what the bill calls “frontier developers”, entities that have trained such a model and had more than $50 million in gross revenue in the preceding year. Smaller players and open-source developers below that threshold are not covered by the transparency and audit requirements in the bill’s current form.
If you’re wondering what the 10^26 number means, that’s the raw compute threshold written into the draft legislation. In plain terms, it’s a measure of how much computational work went into training the model.
To put it in context: 10^26 is 100 septillion operations. That’s an astronomically large number that currently applies only to the largest models being trained by a handful of companies, think the scale of GPT-4, Gemini Ultra, or Claude 3 Opus class models. Most enterprise AI tools, fine-tuned models, and open-source models in common use fall well below that computational number.
What those requirements actually consist of is a combination of transparency disclosures from the developer, third-party audits by licensed “independent verification organizations,” and whistleblower protections for people inside those companies who identify and report safety issues. There’s also a cybersecurity title, Title III, that reauthorizes the Cybersecurity Act of 2015 and includes provisions around securing model weights, data centers, and open-source software. For anyone focused on the supply chain side of AI security, that section is worth reading alongside the Executive Order’s clearinghouse concept, because they are trying to solve adjacent problems from different angles.
The bill’s approach to free speech in Section 141 will generate controversy. It includes language aimed at preventing what it characterizes as AI-driven censorship, which is broadly popular in the abstract and deeply contested in the specifics. That section is likely to be one of the more heavily negotiated parts if the bill moves forward.
What is The CREATOR Act?
The Van Duyne bill is focused exclusively on visual artists and AI-generated stylistic impersonation. It’s not a broad AI regulation bill. It creates a new federal intellectual property right, distinct from copyright, that gives a visual artist the exclusive authority to authorize commercial use of a stylistic impersonation of their work. That right lasts for the life of the artist and up to 50 years post-mortem if renewals are filed.
The bill is careful to carve out general-purpose AI systems. A model that can generate art in a certain style doesn’t automatically create liability just because the capability exists. Liability attaches when a vendor intentionally configures and expressly markets their product to impersonate a specific identified artist commercially. Commentary, criticism, parody, scholarship, and news reporting are all excluded from liability.
For cybersecurity professionals, this bill is mostly background noise, but it matters for anyone advising organizations that use generative AI for marketing, design, or content production. The line it draws between “general-purpose AI that happens to be capable of stylistic mimicry” and “a product marketed for the specific purpose of impersonating Artist X” is going to be litigated hard if this passes.
As a cybersecurity professional, what I’m watching for is whether the June 2nd executive order’s 30-to-60-day implementation deadlines actually produce anything. The CISA Binding Operational Directives, the Treasury-led clearinghouse, and the NSA’s classified frontier model benchmark are all supposed to be in motion by early August. If those land, they’ll tell us whether the Executive Order was a policy statement or a policy. NSPM-11 has its own 90-to-120-day timelines running parallel.
In the meantime, the Great American AI Act’s preemption provision is the most practically significant thing in this entire batch of documents for anyone running an enterprise security or compliance program. If federal law sweeps state AI laws off the table, the fragmented compliance burden that organizations have been anticipating, with different rules in different states for AI deployment, may not materialize in the way people expected. Whether that’s good news or bad news depends heavily on what the federal floor actually looks like when the bill gets finalized, if it does.
These documents represent a significant increase in the pace of AI governance activity at the federal level. None of it is settled. Most of it hasn’t been tested. Read the source documents directly before making any decisions based on them, and track the implementation guidance that follows, because that’s where the real obligations will be defined.
Please keep in mind that the Great American AI Act and the CREATOR Act are discussion drafts as of this writing. Their provisions could change materially before any floor vote. NSPM-11 is in effect as a presidential directive but applies to the national security enterprise, not private organizations. Nothing in this post is legal advice.
I hope you find this post helpful and informative. Thanks for stopping by!
Leave a Reply