AI / Tech · Policy · March 20, 2026

Share

The White House Just Told Congress
How to Regulate AI. Here’s What It Actually Says.

On March 20, 2026, the White House Office of Science and Technology Policy (OSTP) released a 27-recommendation policy document called the National Policy Framework for Artificial Intelligence. OSTP Director Michael Kratsios and then-AI/Crypto Czar David Sacks co-authored it. The document is not a law, not an executive order, and not a regulation. It is a set of legislative recommendations addressed to Congress — a blueprint the administration wants lawmakers to enact before the end of 2026.

The framework spans seven policy areas: child safety, communities, creators, censorship, competitiveness, workforce, and federal preemption of state AI laws. Its central argument is that the United States cannot win the global AI race if 50 states each impose 50 different regulatory regimes on AI developers and deployers. The administration wants Congress to enact a single national standard — and to preempt state AI laws that create what Sacks called “a patchwork difficult for innovators.”

The release coincided with a significant personnel transition: Sacks left his formal role as AI/Crypto Czar after 130 days and moved to co-chair the President’s Council of Advisors on Science and Technology (PCAST) alongside Kratsios — a council whose inaugural membership includes Jensen Huang (Nvidia), Mark Zuckerberg (Meta), Larry Ellison (Oracle), and Sergey Brin (Google). The framework is the administration’s first comprehensive statement of how it believes AI should be governed in America.

• 27legislative recommendationsacross 7 policy areas — child safety, communities, creators, censorship, competitiveness, workforce, and preemption
• EO 14365parent executive ordersigned December 2025 — directed OSTP to prepare this federal AI regulatory framework and evaluate existing state AI laws
• 13PCAST inaugural membersJensen Huang, Mark Zuckerberg, Larry Ellison, Sergey Brin, and 9 others — advisory panel co-chaired by Sacks and Kratsios
• 50-statepatchwork the framework targetsColorado, California, Texas, Utah, and dozens more have enacted AI-specific statutes — framework would preempt those that burden AI development

The framework carries the title National Policy Framework for Artificial Intelligence: Legislative Recommendations. The word “legislative” is load-bearing. This is not a regulation. No agency can enforce it. No court can apply it. It tells Congress what the executive branch wants enacted — and it signals to industry what the administration will support and what it will fight.

The document was required under Executive Order 14365, signed in December 2025, which directed OSTP and the Office of Management and Budget to produce a national AI framework proposal and evaluate existing state AI laws for federal preemption. EO 14365 itself stopped short of preempting any state law — it directed the preparation of recommendations for Congress to do so. This document is that product.

Kratsios framed the goal directly: the administration wants Congress to pass legislation before the end of 2026. House Speaker Mike Johnson (R-LA), Rep. Steve Scalise (R-LA), Rep. Jim Jordan (R-OH), and Sen. Ted Cruz issued statements pledging to work with the framework. Senate Commerce Committee chair Sen. Maria Cantwell signaled openness. On the same day the framework was released, Rep. Donald Beyer (D-VA) and four Democratic colleagues introduced the GUARDRAILS Act to block federal preemption of state AI laws entirely.

“The White House's national AI legislative framework will unleash American ingenuity to win the global AI race, delivering breakthroughs that create jobs, lower costs, and improve lives for Americans across the country.”

Michael Kratsios — OSTP Director · March 20, 2026 · on releasing the framework

The framework’s 27 recommendations are organized under seven headings. Below is the complete structure with every recommendation as drawn from the primary document and corroborated across multiple law-firm analyses of the official PDF.

The preemption section is the most consequential and contested part of the framework. As of March 2026, more than a dozen states have enacted AI-specific statutes: Colorado has algorithmic accountability requirements, California has passed multiple AI bills covering hiring, healthcare, and deepfakes, Texas has its own framework, and Utah has enacted AI disclosure rules. The framework treats this legal landscape as a problem to be solved, not a feature to be preserved.

The preemption proposal is not a blanket override of all state authority. The framework explicitly carves out state enforcement of generally applicable consumer protection laws, child safety statutes, government procurement rules, and zoning authority. What it targets are AI-specific state laws that go beyond those generally applicable frameworks — laws that, in the administration’s view, “ban particular verticals” or “impose undue burdens on AI development.”

“We believe that we can't have preemption without other stuff attached to it, and it has to be a give and take with both.”

Michael Kratsios — OSTP Director · Nextgov/FCW interview · March 2026

The framework was released on the same week David Sacks formally concluded his 130-day tenure as the White House AI and Crypto Czar. Sacks did not leave; he changed roles. President Trump appointed him co-chair of the President’s Council of Advisors on Science and Technology (PCAST), alongside Kratsios. The significance of the transition is structural: as AI Czar, Sacks held a formal White House staff role with direct operational authority. PCAST is an external advisory body — it studies issues, makes recommendations, and advises the president, but does not issue orders or direct agencies.

Sacks described the expanded scope himself: as PCAST co-chair, he “can now make recommendations on not just AI but an expanded range of technology topics.” The PCAST mandate covers AI, semiconductors, quantum computing, and nuclear power. The council’s near-term priority, confirmed by TechCrunch, is advancing the national AI framework just released through Congress. The roster of 13 inaugural PCAST members — including Jensen Huang, Zuckerberg, Ellison, Brin, and AMD CEO Lisa Su — gives the advisory body unusual gravitational pull for what is technically a non-governmental advisory role.

The framework’s creators section contains the most politically unusual position in the document: the administration explicitly states it believes AI training on copyrighted material “does not violate copyright laws,”but simultaneously recommends that Congress not legislate that conclusion — instead allowing courts to settle the question through litigation. The reasoning is that premature legislation could “impact the fair use judicial resolution,” which the administration appears to want resolved in industry’s favor by precedent rather than statute.

The framework’s compensatory mechanism for creators is the voluntary collective licensing framework— a proposal allowing creators to pool bargaining power and negotiate compensation from AI providers without triggering antitrust liability. Whether voluntary frameworks will attract the major AI labs, who benefit from the status quo, is an open question the framework does not answer. On digital replicas, the framework backs statutory protection against unauthorized use of a person’s voice, likeness, or image — an area where several states have already acted and the federal gap is widely acknowledged.

The framework does not name the EU AI Act explicitly, but the contrast is constant and deliberate. The EU AI Act, which entered force in August 2024 and began full enforcement in August 2026, classifies AI applications into risk tiers (unacceptable, high, limited, minimal) and imposes compliance requirements by risk level — using what Kratsios described at CSIS as “antiquated high-risk/low-risk categories.” The EU approach creates a horizontal AI regulator. The U.S. framework explicitly recommends not creating a new federal AI regulatory body — relying instead on existing sector-specific regulators (the FDA for medical AI, the DOT for autonomous vehicles, the FTC for consumer protection).

Republican leaders in both chambers issued statements of support within hours of the framework’s release. House Speaker Johnson and Rep. Babin (House Science Committee) pledged to work with OSTP on legislation. Sen. Cruz, chair of the Commerce Committee’s relevant subcommittee, aligned with the preemption posture. The reception on the Republican side was not universal: Sen. Marsha Blackburn (R-TN) has circulated a competing draft with stricter child protections and developer accountability standards that exceed what the White House framework recommends.

Democratic opposition coalesced around preserving state authority. On March 20, Rep. Beyer (D-VA), Rep. Doris Matsui (D-CA), Rep. Ted Lieu (D-CA), Rep. Sara Jacobs (D-CA), and Rep. April McClain Delaney (D-MD) introduced the Guaranteeing and Upholding Americans’ Right to Decide Responsible AI Laws and Standards (GUARDRAILS) Act. The GUARDRAILS Act would repeal EO 14365 and effectively block any federal moratorium on state AI regulation. Passage of the GUARDRAILS Act in a Republican-controlled Congress is not expected.

The Lawfare analysis of the framework catalogues its coverage gaps, and they are significant. The 27 recommendations address seven of the most politically salient areas — but they omit whole domains that courts, regulators, and technologists treat as central to AI governance:

The gaps are not accidental. The administration’s stated preference is to leave algorithmic accountability and high-stakes AI deployment regulation to existing sector regulators — the FDA for medical AI, the CFPB for credit AI — rather than creating a new federal AI agency or a horizontal compliance requirement. Critics argue this leaves exactly the domains where existing regulators lack AI expertise most exposed to harms without specific safeguards.