Understanding America’s AI Action Plan, Part II

The White House asks Congress to make it permanent

Last July, in Understanding America’s AI Action Plan, I walked through the Trump Administration’s 23-page executive blueprint for achieving what the White House called “a new golden age of human flourishing” through AI. The Plan was ambitious, deregulatory, and unmistakably “AI Evangelist” in its posture. I praised some of it, criticized some of it, and concluded with the blunt assessment that, good plan or bad plan, it was The Plan. There was no Plan B.

It’s eight months later, and there is still no Plan B. But there is a Plan A, Part II!

Today, March 20, 2026, the White House released a four-page document titled National Policy Framework for Artificial Intelligence: Legislative Recommendations. Where the original AI Action Plan was an executive directive, aimed at federal agencies and signed by the President’s own hand, this new document is aimed at Congress. It asks the legislature to codify the Administration’s AI vision into statute. Executive orders, after all, are written in sand. What one President decrees, the next can erase. If President Trump wants his AI policy to survive a change in administration, he needs legislation.

The document is organized into seven pillars. Each deserves scrutiny. Let’s dive in.

I. Protecting Children and Empowering Parents

The first pillar calls on Congress to protect children from AI-related harms. It invokes the Take It Down Act, a signature initiative of First Lady Melania Trump that targets deepfake abuse of minors. It then calls for age-assurance requirements, parental controls over privacy settings and content exposure, and features to reduce the risks of sexual exploitation and self-harm.

This is uncontroversial in the way that “protecting children” is always uncontroversial. Nobody campaigns on a platform of not protecting children. The question, as always, is what “protection” means in practice and who bears the cost.

The document does include one genuinely important caveat: “Congress should avoid setting ambiguous standards about permissible content, or open-ended liability, that could give rise to excessive litigation.” That’s a nod to the reality that vague “child safety” mandates tend to become censorship tools. The EU’s approach to online child safety has already demonstrated the ratchet: once you establish a framework for restricting content “for the children,” the definition of harmful content expands to include whatever the regime finds inconvenient. The White House is signaling that it doesn’t want that.

It also preserves state authority to enforce “generally applicable laws protecting children, such as prohibitions on child sexual abuse material, even where such material is generated by AI.” This is notable because it means AI-generated CSAM is to be treated, legally, the same as the real thing. That’s a defensible position, but it does raise interesting questions about where the line sits for other forms of AI-generated content that depicts no real person doing anything real.

I’ll return to this tension later, because it resurfaces in a more dangerous form in Pillar VII.

II. Safeguarding and Strengthening American Communities

The second pillar is a grab bag. It bundles together the energy grid, consumer protection, national security, and small business support into a single section that reads like it was drafted by a committee, probably because it was.

Three items stand out.

First, the Ratepayer Protection Pledge. Congress is asked to “ensure that residential ratepayers do not experience increased electricity costs as a result of new AI data center construction and operation.” Recall that in my July essay, I noted the AI Action Plan was the obituary of the Green Energy movement. This legislative recommendation confirms it. The White House is telling the energy industry: build as fast as you can, but don’t stick grandma with the bill. The solution is to “streamline federal permitting” so AI developers can build “on-site and behind-the-meter power generation.” In plain English: let the tech companies build their own power plants next to their data centers, bypassing the grid entirely.

This is actually clever. It means that AI infrastructure buildout doesn’t compete with residential power consumption, because the data centers generate their own electricity. It also means that when the White House says “embrace new energy generation sources at the technological frontier (e.g. enhanced geothermal, nuclear fission, and nuclear fusion),” the primary customer for those frontier energy technologies will be the AI companies themselves. The Department of Energy becomes, in practice, a support apparatus for Silicon Valley’s power needs.

Second, the document calls for Congress to “ensure that the appropriate agencies within the national security enterprise possess sufficient technical capacity to understand frontier AI model capabilities.” Translation: the intelligence community needs its own AI expertise so it isn’t wholly dependent on OpenAI and Google to explain what their own models can do. This is the bureaucratic version of “trust but verify.” It is also an acknowledgment that we are building weapons we do not fully understand. I noted this problem in my discussion of AI interpretability last July, where the entire challenge was dismissed in a single bullet point. Here it surfaces again, wearing a suit and speaking Pentagonese.

Third, the call to “augment existing law enforcement efforts to combat AI-enabled impersonation scams and fraud that target vulnerable populations such as seniors.” This is the first hint that AI is already causing real harm in the real world, outside the seminar room debates about alignment and existential risk. Grandma isn’t worried about paperclip maximizers. She’s worried about the phone call that sounds exactly like her grandson asking for bail money.

III. Respecting Intellectual Property Rights and Supporting Creators

Here is where the document gets genuinely interesting, because here is where the White House stakes out a position on the most contentious legal question in AI: whether training on copyrighted material constitutes fair use.

The answer is… a masterpiece of strategic ambiguity:

Although the Administration believes that training of AI models on copyrighted material does not violate copyright laws, it acknowledges arguments to the contrary exist and therefore supports allowing the Courts to resolve this issue.

Read that again. The White House has just told you what it thinks the law is. It has also told you it will not lift a finger to make its view binding. Instead, it punts to the judiciary.

This is, in the language of my prior legal analyses, a plausible but non-binding signal. The Administration is putting its thumb on the scale without actually touching the scale. If you’re an AI company, you read this and feel reassured. If you’re a content creator, you read this and feel abandoned. If you’re a federal judge, you read this and shrug, because the executive’s opinion on fair use is worth precisely nothing in an Article III courtroom.

The more substantive recommendation is the call for Congress to “consider enabling licensing frameworks or collective rights systems for rights holders to collectively negotiate compensation from AI providers, without incurring antitrust liability.” This is significant. Under current law, if every photographer in America got together and said “We’re collectively demanding $X per image from OpenAI,” that would be a textbook antitrust violation, a price-fixing cartel. The White House is suggesting Congress create a specific exemption so rights holders can bargain collectively against the tech giants.

But note the carefully embedded poison pill: “Any such legislation, however, should not address when or whether such licensing is required.” Congress can build the negotiating table, but it cannot force anyone to sit down at it. The AI companies retain the right to argue that they owe nothing. The licensing framework is a gesture toward fairness that may, in practice, change nothing.

The digital replica provision is more straightforward: a federal right of publicity that protects your voice, likeness, and “other identifiable attributes” from unauthorized AI replication, with carve-outs for parody, satire, and news reporting. This is the deepfake defense for adults, mirroring the child-protection provisions in Pillar I. If someone uses AI to make a video of you saying things you never said, you’ll have a federal cause of action. Unless it’s funny, newsworthy, or political, in which case the First Amendment eats the statute. Maybe.

IV. Preventing Censorship and Protecting Free Speech

Pillar IV is the shortest section and, for that reason, potentially the most revealing.

The White House calls on Congress to “prevent the United States government from coercing technology providers, including AI providers, to ban, compel, or alter content based on partisan or ideological agendas.” It further asks for “an effective means for Americans to seek redress from the Federal Government for agency efforts to censor expression on AI platforms.”

This is, unmistakably, a response to the censorship-industrial complex that operated during the Biden Administration, when federal agencies coordinated with social media companies to suppress speech under the banner of fighting “misinformation.” The White House wants Congress to make that kind of operation illegal, and to give citizens a way to sue the government when it happens.

As a matter of principle, I applaud this entirely.

As a matter of practice, I have the same objection I raised in July. The Left’s commitment to free speech is always provisional. When the Left next holds the White House, “preventing government coercion of AI providers” will be reinterpreted, narrowed, or simply ignored in favor of “preventing harmful content.” The statutory language will matter enormously. A well-drafted statute might survive such reinterpretation. A poorly drafted one will become tissue paper.

What is conspicuously absent from this section is any mention of the AI companies’ own ideological biases. The July Action Plan at least addressed this, calling for procurement guidelines that ensure frontier LLMs are “objective and free from top-down ideological bias.” The legislative recommendations say nothing about it. The White House has retreated from even the modest ambition of using federal purchasing power to pressure the labs into neutrality. Whether this reflects a genuine policy shift or merely an acknowledgment that Congress won’t legislate AI objectivity, the effect is the same: the bias problem is now entirely unaddressed in the proposed statutory framework.

I said in July that what we need is “real transparency into design and training in conjunction with a plurality of options available to match the plurality of our ideologies.” The legislative recommendations move us no closer to that goal.

V. Enabling Innovation and Ensuring American AI Dominance

Pillar V is the most predictable section. It’s got regulatory sandboxes and open federal datasets but no new federal AI regulatory body. It asks existing regulators to handle AI in their own sectors.

The “no new regulatory body” position is the most consequential item here, and it cuts both ways. On one hand, creating an “AI Commission” or “Federal AI Authority” would invite regulatory capture, bureaucratic empire-building, and the inevitable drift toward suppressing innovation in the name of safety. Every new agency becomes a jobs program for the credentialed class that then fights to justify its own existence. The White House is right to be wary.

On the other hand, the absence of a dedicated regulator means that AI governance will be fragmented across dozens of agencies, none of which has AI as its primary mission, most of which lack the technical capacity to understand what they’re regulating. The SEC will regulate AI in securities. The FDA will regulate AI in medicine. The FCC will regulate AI in communications. Each will develop its own standards, its own enforcement posture, its own body of precedent. The result will be a patchwork, not a framework.

The White House is betting that a patchwork of sector-specific regulation is better than a monolithic AI regulator. That bet may be correct. But it should be recognized as a bet, not a certainty. It is, for whatever it’s worth, the opposite of what we did with atomic power.

VI. Educating Americans and Developing an AI-Ready Workforce

If the workforce section of the July Action Plan was “almost dark comedy,” this legislative version is the joke stripped of its punchline.

Congress is asked to use “non-regulatory methods” to incorporate AI training into existing education and workforce programs. It is asked to “expand Federal efforts to study trends in task-level workforce realignment driven by AI.” It is asked to “bolster capabilities at land-grant institutions” to develop AI youth programs.

Study trends? Pilot programs? Land-grant institutions? Bluntly, this is the vocabulary of a government that has no idea what to do and is trying to buy time by studying the problem.

I said in July that I didn’t have a better plan. I still don’t. But the gap between the rhetoric (”AI will transform how work gets done across all industries and occupations”) and the response (”let’s study it at land-grant colleges”) has grown from comedy to something more unsettling. It has been eight months since the original Action Plan. In those eight months, AI capabilities have advanced substantially. Claude, GPT, Gemini, and their kin can now do things that were speculative when the first Plan was written.

The Administration still takes the position that AI will change everything about the economy except the need for Americans to work. Perhaps it has to take that position. A sitting President cannot tell the electorate, “Your jobs are going away and we don’t know what comes next.” But the silence on post-labor economics, on alternative arrangements, on what happens when the tractors actually arrive, is deafening.

America’s draft horses can see the tractors on the horizon, and the government’s response is essentially to fund a study on equine career counseling.

VII. Establishing a Federal Policy Framework, Preempting Cumbersome State AI Laws

The seventh and final pillar is, in my assessment, the most important section in the document, and the one that deserves the most careful reading.

The White House asks Congress to “preempt state AI laws that impose undue burdens to ensure a minimally burdensome national standard.” It then carves out three areas where states retain authority: traditional police powers (child protection, fraud, consumer protection), zoning laws for AI infrastructure, and requirements governing a state’s own use of AI.

Then comes the key language:

States should not be permitted to regulate AI development, because it is an inherently interstate phenomenon with key foreign policy and national security implications.

States should not unduly burden Americans’ use of AI for activity that would be lawful if performed without AI.

States should not be permitted to penalize AI developers for a third party’s unlawful conduct involving their models.

Take these in order.

The first principle, that states cannot regulate AI development, is a powerful assertion of federal supremacy. It means that if California, or Texas, or any state passes a law imposing safety requirements, transparency mandates, or testing obligations on AI developers, Congress can sweep it aside. The justification is that AI development is “inherently interstate,” which is true, and has “national security implications,” which is also true, but those same arguments could justify federal preemption of state regulation in virtually any technology sector. The interstate commerce clause has been stretched to cover wheat grown for personal consumption; it can certainly cover code compiled in San Francisco and deployed in Topeka.

The practical effect is to concentrate AI governance in Washington, D.C., where the tech lobby is strongest and the regulatory posture is lightest. States like California and New York that have attempted to impose AI safety requirements will see those efforts nullified. Given that New York’s proposed regulation was a regulatory power grab by the professional cartels, I’m happy enough with that outcome.

The second principle, that AI use should not be burdened for lawful activity, sounds anodyne until you think about it. If it is lawful for a human to write a political advertisement, it is lawful for an AI to write one. If it is lawful for a human to apply for a job, it is lawful for an AI to submit the application. If it is lawful for a human to practice medicine (with a license), then... what? The principle, taken to its logical end, means that any regulation of AI-assisted activity must be justified on grounds other than “an AI did it.” The AI itself is a legally neutral instrument, like a hammer. You can regulate hammering your neighbor, but you cannot regulate hammers. You can’t “burden” the AI’s activity.

There is a coherent libertarian case for this position. I am highly sympathetic to it! But it does mean that disclosure requirements, for instance requiring that AI-generated content be labeled as such, would likely fail under this framework. If a human can publish an unlabeled opinion piece, an AI-generated opinion piece needs no label either. This has implications for the information environment that the White House’s free speech section does not address. Perhaps it’s unintended, or perhaps it’s specifically intended under an expectation that most work is going to be done by AIs. Or perhaps the idea is that the disclosure requirements will end up enforced by third-party terms of service, like YouTube requiring a disclosure if content was made with AI. Letting the market set the disclosure standards could be viable, at least to the extent that the markets aren’t oligopolies in bed with government.

The third principle, that developers cannot be penalized for third-party misuse, is the most legally consequential. It is, in effect, Section 230 for AI. Just as social media platforms are not liable for user-generated content, AI developers would not be liable for what users do with their models. If someone uses an open-weight model to generate a bioweapon schematic or a deepfake of a senator, the developer bears no legal responsibility.

Recall that in July I discussed the open-weight question at length, noting Geoffrey Hinton’s warning that releasing these models is “like handing out blueprints for nuclear weapons.”¹ The White House’s legislative recommendation now asks Congress to codify the principle that the blueprint distributor bears no liability for what is built from the blueprint. This is a remarkable position for an Administration that simultaneously calls for AI safety and national security readiness.

Obviously, we could argue that liability should attach at the point of misuse, not at the point of development, just as gun manufacturers are generally not liable for shootings. But the analogy is imperfect. A gun is a physical object sold through a regulated supply chain. An open-weight model is a digital artifact that can be downloaded, copied, modified, and deployed by anyone with an internet connection and sufficient compute. The supply chain doesn’t exist. There is no point of sale, no background check, no serial number. A 3D printed gun blueprint is perhaps a closer analog, or the DNA of a pathogen are closer. I’m willing to tolerate 3D printed gun blueprints in a free society; I’m not so sure I’d tolerate Spanish Flu blueprints.

My suspicion of centralized governmental and corporate power (what I call “Tyranny, Inc.”) makes me favor open-weight models in private hands. I want people to be able to ask AI for help with medicine, law, and other matters that are currently controlled by regulatory cartels. I myself have set up a private open-weight model for my family’s use, and at some point we will likely use Pliny the Liberator’s Obliteratus protocol to remove the unwanted safety guardrails that constantly restrict discussions of, e.g. health care. For that possibility to exist, there needs to be sort of liability shield for the AI companies. But it is a certainty that there will be people who use AI to harm themselves (out of stupidity) and harm others (out of malice), possibly at great or even breathtaking scale.

I genuinely don’t know the right answer here. I don’t think anyone does. How do you build long-term policy for a technology whose agentic capabilities are doubling every 3 months? Do we just pray for the S curve to flatten?

This section, more than any other, reveals the fundamental tension at the heart of American AI policy: the desire for maximum innovation and minimum accountability is currently held together by the hope that the benefits will outweigh the harms.

The Missing Sections

Reading the legislative recommendations alongside the July Action Plan, it’s striking to notice how much of the Action Plan didn’t make it in. The July Plan addressed open-source and open-weight models at length. The legislative recommendations say almost nothing about them. Either the White House has decided this is better handled through executive action, or it has concluded that Congress will not legislate on the subject. Either way, the open-weight question, one of the most consequential in AI policy, has been left to the market.

The July Plan discussed AI interpretability, albeit in a single bullet point. The legislative recommendations do not mention it at all. The black box remains black. The July Plan also addressed the energy crisis in detail, warning of a “confluence of challenges” that demanded “strategic foresight and decisive action.” The legislative recommendations reduce this to a single bullet about ratepayer protection. The energy problem hasn’t gone away but it’s been moved off the legislative agenda.

And the July Plan, however inadequately, grappled with the question of AI bias and ideological neutrality. The legislative recommendations have abandoned even the pretense of addressing it.

There Are Still No Brakes On The AI Train, And The Track is Now Being Laid

What we are witnessing is the transformation of an executive aspiration into a legislative program. The broad vision of the July Action Plan is being narrowed, sharpened, and codified into the specific statutory changes that the White House believes it can actually get through Congress.

In that narrowing, certain priorities have survived: deregulation, federal preemption, innovation acceleration, child protection, free speech. These are the items the Administration believes are politically achievable.

Other priorities have been dropped: AI interpretability, open-weight governance, energy infrastructure, ideological bias. These are the items the Administration has decided are either too technically complex, too politically toxic, or too difficult to legislate. Or perhaps they’re just a bit distracted by foreign affairs and forgot to include them. There’s been a lot going on, man.

In any case, the result is a legislative framework optimized for one thing above all else: speed. Remove the regulatory barriers. Preempt the states. Protect the developers from liability. Let the AI companies build. Build fast. Build now. Build before China does.

This is not an unreasonable posture for a government that believes it is in a civilizational race. As I wrote in Build AI or Be Buried By Those Who Do, every faction of the American elite has converged on AI development because they see no alternative.

But speed has costs. What is built fast is not always built well. What is built without accountability is not always built safely. And what is built to beat China might break America. Ask your preferred LLM to contemplate this for you on the Tree of Woe.

1. I further sketched out these problems in the extended analogy presented in my article The West’s Warhammer Moment. Unfortunately the analogy was a little too extended, and while we had a good discussion of Warhammer 40K we didn’t have a very good discussion of AI.