Without a Federal AI Law, How Is the United States Rewriting Global AI Rules?

By Zhiqian Wu

Introduction: An AI Superpower Without a Federal AI Law

The United States has the world's strongest AI companies, the most computing power, and the most active capital markets, yet it still has no unified federal AI law.

As of 2026, the main force regulating AI is not Washington but the states. In 2025 alone, 38 states passed 145 AI-related laws. At the same time, the White House has been pushing for federal law to "preempt" state law at home, while steadily tightening chip and technology exports abroad.

This is the basic picture of U.S. AI governance: not one set of rules, but several sets pulling against each other. This article follows three lines of divergence (federal vs. state, state vs. state, and domestic vs. foreign) and asks how this divergence spills over to shape AI development around the world.

I. Federal vs. State: A Tug-of-War Lasting More Than a Year

The core question is simple: who has the authority to set the rules for AI? Over the past year and more, the White House has tried three times to take that authority back, yet state laws remain fully intact.

Round one came in July 2025. Congress tried to add a moratorium on state AI legislation to a budget bill, but the Senate struck it down by a vote of 99 to 1.

Round two was Executive Order 14365 in December 2025. It created a Department of Justice AI Litigation Task Force to challenge state AI laws, and tied federal broadband funding under the BEAD program to whether states abandon "onerous" AI rules. But an executive order cannot repeal state laws on its own.

Round three came in March 2026, when the White House released its National AI Policy Framework, recommending that Congress pass legislation allowing federal law to preempt state AI laws that impose an "undue burden." The framework is a non-binding recommendation. By September 2026, a bipartisan preemption draft had been stalled for months, the Speaker of the House had publicly questioned whether Congress could "lead" on AI safety, and the legislative outlook remained unclear.

The federal argument is about competitiveness: 50 different sets of state rules drive up compliance costs, slow innovation, and could cause the United States to lose ground in its competition with China. The White House says it wants a "minimally burdensome" national standard rather than 50 conflicting state standards.

The states' argument is about protection: with Washington slow to legislate, consumer protection, anti-discrimination, and child safety cannot wait. Notably, the March framework itself acknowledges the need to preserve states' traditional powers over child protection, fraud, and consumer safety.

The most interesting result is a "the harder you push, the more you get" effect. Signals of preemption did not scare off state legislatures; instead, they sped up lawmaking, as states rushed to put their own frameworks in place before any federal law arrived. In the first quarter of 2026 alone, about 25 new state AI laws were added. Opposition to preemption doesn't come only from Democrats, either: the 99-to-1 vote in 2025 showed that there are plenty of "states' rights" Republicans as well.

II. State vs. State: One Country, Different Paths for AI

Even at the state level, there is no single answer. California chose to focus on the largest models, while Colorado retreated from having the toughest law in the country to "disclosure only."

California's SB 53 took effect on January 1, 2026, and covers developers of frontier models trained with more than 10^26 floating-point operations (FLOPs). Another law, SB 942, took effect on August 2, 2026, and requires generative AI systems with more than 1 million monthly users to label AI-generated content. Colorado, meanwhile, replaced its 2024 law with a new one, SB 26-189, in May 2026, shifting the focus from risk management and impact assessments to information disclosure and pushing the effective date back to January 1, 2027.

California: the frontier model transparency path. SB 53 is the first state law in the U.S. to directly regulate frontier AI developers. Its predecessor, SB 1047, was vetoed by the governor in 2024; SB 53 is a narrower, more transparency-focused replacement. Because most major AI companies are headquartered in California, this state law in practice carries national and even global influence.

Colorado: from "toughest" to "retreat." Passed in 2024, Colorado's AI law was once considered the most comprehensive state AI law in the United States, requiring companies to exercise a duty of care over high-risk AI and conduct impact assessments. But its effective date was repeatedly postponed. In April 2026, xAI sued the state on First Amendment, equal protection, and dormant Commerce Clause grounds. In May, the governor signed a new law repealing the original and replacing it with lighter disclosure rules. The original law never took effect for a single day.

Colorado's story is almost a textbook case of the tension between innovation and regulation: corporate litigation, federal pressure, and the state's own concerns about its industries combined to push a state into changing course on its own.

III. Loosening at Home, Tightening Abroad

The third line of divergence is the easiest to overlook: the United States advocates "light-touch regulation" for domestic AI development while steadily ramping up controls abroad. The two stances are not contradictory. They follow the same logic: help American AI run faster, while making competitors run slower.

Chip controls: rules that keep changing. In 2025, the "AI Diffusion Rule" left behind by the Biden administration, which restricted chip exports by sorting countries into three tiers, was rescinded by the Trump administration. A new rule in January 2026 instead set performance thresholds, capped sales to China at no more than 50% of shipments to the U.S., and required end-use certification. That same month, a 25% tariff was added to AI chip exports to China, layering trade policy on top of national security licensing.

Controls are extending to the model layer. In June 2026, access to two Anthropic models, Claude Fable 5 and Mythos 5, was suspended to comply with U.S. Department of Commerce export controls; access was restored on July 1 after the controls were lifted on June 30. This shows that export controls are no longer aimed only at hardware: frontier models themselves are now coming into view.

Mostly unilateral action. Under the Wassenaar Arrangement, which involves 42 countries, members have yet to reach consensus on binding AI chip controls. The United States has largely moved on its own and then asked allies to follow.

"Exporting the American AI stack." Alongside "blocking" comes "pushing": U.S. policy encourages exporting "full-stack AI packages" of hardware, models, software, applications, and standards to allies. In other words, the United States wants other countries to use American AI, rather than build their own or use someone else's.

IV. Spillover: How U.S. Divergence Shapes Global AI

America's internal divisions do not stay inside America. They affect the world through at least four channels.

1. Regulatory competition: are global rules loosening overall? The European Union was once seen as the benchmark for AI regulation, but it too is pulling back. On July 27, 2026, the EU's Digital Omnibus formally took effect, postponing obligations for high-risk systems under the AI Act to December 2027 and August 2028. The official rationale was to simplify rules, reduce compliance costs, and create a more innovation-friendly environment. America's light-touch stance is likely one source of the competitive pressure the EU has felt. Critics, however, argue that the delay has not brought certainty and has instead created new uncertainty for businesses.

2. Multinational companies: complying across multiple rulebooks. A global AI company must deal with rules from multiple U.S. states, the EU, and China at the same time. One possibility is that companies simply comply with the "strictest standard" across the board, making California's or the EU's rules the de facto global standard. Another is that companies exploit regulatory differences, placing sensitive operations wherever oversight is lightest.

3. The Global South: a widening compute gap. The combination of chip controls and the "full-stack export" strategy means that developing countries' access to advanced computing power increasingly depends on their relationship with the United States. For countries without their own chips or data centers, AI development may become not just a technical question but a matter of diplomatic and geopolitical choice. This point is often overlooked in discussions, yet it could profoundly shape the global development landscape over the next decade.

4. Multilateral governance: no longer one American voice. When Washington and Sacramento (California's capital) give different answers to the same question, it becomes hard for others to tell what the "American position" actually is. This weakens the consistency of the U.S. voice on multilateral AI governance platforms such as the United Nations and the G7, and leaves room for other countries to put forward their own governance proposals.

V. What This Means for Chinese Readers

For companies going global: compliance must reach down to the state level. Entering the U.S. market means looking beyond federal policy. Until federal preemption is settled, state laws remain valid and enforceable, and companies need to assess their obligations state by state.

For industry: external constraints are a long-term variable. America's combination of accelerating at home and restricting abroad is unlikely to change in the short term. Access to computing power, access to models, and technology partnerships should all be planned on the assumption that the rules will keep changing.

For governance thinking: both models have costs. America's decentralized governance lets states experiment, but it also brings fragmentation and uncertainty. Centralized, unified regulation is clearer and faster to enforce, but may leave less room for trial and error with new technologies. There is no simple answer to which is better; what matters is finding a sustainable balance between protection and innovation.

Conclusion

The tension between innovation and regulation is not unique to the United States. What makes the U.S. case distinctive is that this tension plays out on three levels at once: between the federal government and the states, among the states themselves, and between domestic and foreign policy.

Over the next year or two, three things are worth watching: whether Congress can pass federal preemption legislation; how state laws such as California's SB 53 fare in enforcement and litigation; and whether U.S. export controls expand further from chips to models. The answers to these three questions will go a long way toward shaping the map of global AI rules.

References

• The Federal AI Moratorium and State-Preemption Fight (CASRAI)

• State AI Laws Take Hold as Federal Preemption Stalls (Cloud Security Alliance)

• President Trump Unveils AI EO Advancing Federal Preemption (Akin Gump)

• White House National AI Framework and State Preemption (AlgeriaTech)

• US AI Enforcement Accelerates as Federal Policy Stalls (AlgeriaTech)

• 2026 AI Policy and Semiconductor Outlook (ML Strategies)

• State AI Laws in 2026: SB 53, SB 942, and Preemption (GC AI)

• State AI Laws by State (Layer3 Labs)

• AI Chip Export Controls in 2026 (CASRAI)

• Compute Governance: AI Chips Export Controls (Longterm Wiki)

• EU AI Act Digital Omnibus Deadline Delay (CASRAI)

• EU's Digital Omnibus Proposes Major Delay (CADE Project)

• Anthropic statement on access to Fable and Mythos