Who Will Actually Regulate American AI?
Democrats are projected to have won both chambers of Congress in the 2026 midterms, and President Trump has already signaled how he intends to meet them. “They’ll send over things that are insane and I’ll veto them,” he said in a Fox News interview after the election (November 4, 2026), days after embracing the “Mr. Veto” label on social media. His theory is that a few months of vetoes will bring Democrats to the table for deals. Whether that happens is an open question. What matters for this argument is the environment taking shape: divided government, a president with a veto and an executive pen, and a Congress that has never passed a comprehensive AI framework. The election results are still projections, but the institutional problem does not depend on the final certification.
This is not a forecast of instability. It is the operating environment, and it is one in which coherent federal AI governance is least likely, though not impossible. The United States remains the world’s leading AI developer, but development and deployment are governed by different logics. Capital and talent follow capability; deployment follows rules, or the absence of them. What the midterms have produced is not a regulatory vacuum but a default regime: executive orders that swing with each administration, state laws that can become national product requirements by accident, federal litigation against those same laws, and European enforcement that reaches American firms regardless of what Washington does.
The question is no longer whether American AI will be regulated. It is who writes the rules by default, and who can be held accountable for the results. The answer depends on the layer. For frontier-model developers that ship globally, Brussels and a few active states can set the practical compliance floor, while federal litigation contests how long some state rules last. For deployers of AI in hiring, benefits, insurance, and healthcare, state legislatures and attorneys general do much of the work. In Washington, the executive branch supplies direction but not durability. Above all three layers, courts determine the legal boundaries within which the others operate. The problem is not that no one is accountable for anything. It is that authority and accountability are dispersed across institutions that do not share a single mandate or answer for the system as a whole.
Why Washington is unlikely to write the rulebook
The obstacle is not simply that Democrats and Republicans disagree. Under current Senate rules, a comprehensive statute would ordinarily need to clear a 60-vote threshold to overcome a filibuster, followed by presidential approval, or a two-thirds majority in both chambers to override a veto. That is a difficult path in a period when majorities change hands every few years and neither party can count on a durable governing coalition. Procedural exceptions, including reconciliation, exist, but they do not remove the underlying political problem.
Narrow agreement is possible, and it is worth being precise about its limits. Congress passed the TAKE IT DOWN Act in 2025, which targets nonconsensual intimate imagery, including AI-generated imagery. The Senate voted 99–1 that same year to strip a blanket moratorium on state AI laws from a budget bill. The CHIPS and Science Act of 2022 shows that bipartisan industrial policy can survive. But these examples share a feature: they address a specific harm, reject a specific proposal, or distribute benefits rather than impose a comprehensive set of duties. None is a rulebook for how AI may be built and deployed.
The weakening of party attachments sharpens the institutional problem. Gallup’s 2025 data show a record 45% of Americans identifying as political independents, a shift that has proven sticky across generations rather than a temporary phase of youth. But this is not a moderation story. Most of these independents still lean toward and vote reliably for one party, rejecting the label rather than the two-party duopoly. Furthermore, the surge partly reflects transient discontent with whichever party holds power.
The analytical point is not that independent voters are driving legislative chaos, but that looser party attachments are a symptom of a highly volatile political environment. Combined with polarization, primary-election incentives, and geographic sorting, these conditions make durable federal statutes structurally harder to build. A framework passed by one coalition can invite revision or reversal by the next. That, more than any single polling number, is what the default regime has to work around.
Three layers, one pattern
The federal layer: direction without durability
The Trump administration revoked Biden-era AI safety requirements. On December 11, 2025, Trump signed an executive order directing the Attorney General to challenge state AI laws as preempted or unlawful under a “minimally burdensome national policy framework.” Two features of that order matter. First, an executive order cannot itself preempt state law; it directs litigation, and courts decide the outcome. Second, the framework is an executive creation, revocable by the next president, so for labs and deployers the planning horizon may be no longer than the current administration.
The federal posture is also hostile to what the White House calls “overregulation by the States.” That limits how strict any national baseline is likely to become under this administration. The likely equilibrium is therefore neither heavy regulation nor none. It is incoherent regulation: strict in some states, comparatively hands-off federally, and repeatedly litigated.
The frontier-model layer: contested and fast-moving
California’s Transparency in Frontier AI Act and Texas’s Responsible AI Governance Act took effect on January 1, 2026. Companies with nationally distributed products have an incentive to build around the strictest applicable state requirements rather than maintain separate compliance architectures for every jurisdiction. In practice, that can turn state legislation into national product requirements. It is not automatic, however. Firms can geographically restrict products, maintain regional versions, or decide that a market is not worth serving. The extent of California’s influence will depend on those commercial choices as well as the law itself. Meanwhile, these laws operate under a federal litigation cloud, and the record so far shows how unsettled the outcome is.
The best illustration is Colorado. On April 9, 2026, xAI sued to block the state’s comprehensive AI Act before its June 30 effective date. On April 24, the Justice Department intervened, arguing, among other things, that the law violates the Equal Protection Clause. On April 27, a federal court stayed enforcement. Colorado later repealed and replaced the act with a narrower automated decision-making law scheduled to take effect on January 1, 2027. Note the mechanism: a private plaintiff sued, and the federal government joined. That is not a top-down preemption campaign. And the courts have not moved in one direction. In xAI’s separate challenge to California’s training-data transparency law, a district court declined to issue a preliminary injunction, and the case went to the Ninth Circuit. A state law can remain in force while under federal legal attack, or be stayed before it ever takes effect. Companies must account for the rules currently in force while planning for the possibility that a court may limit or invalidate them.
Brussels is the other force on this layer. Since August 2, 2026, the European Commission has held live enforcement powers over general-purpose AI models. It can run evaluations, demand corrective measures, restrict EU market access, and fine providers up to €15 million or 3% of global annual turnover, whichever is higher. Refusing an information request is itself finable. Sidley Austin partner Elisabetta Righini told CNBC that a US address does not put a lab outside the regulator’s reach, and the Act’s territorial scope turns on where a model is placed on the market, not where its developer is headquartered. The practical effect is slower and narrower than the formal powers suggest: models placed on the market before August 2025 have until August 2027 to comply, and the AI Office issued its first formal Requests for Information on August 29, 2026; market restrictions or fines had not been imposed as of this writing.
The EU rules therefore do not automatically become a global standard. Their influence depends on whether firms find it more economical to maintain one product and compliance architecture or to create regional versions. For companies unwilling or unable to fork their products by region, the EU requirements can become a de facto global design constraint. At minimum, they establish the rules for access to a major market, and those rules do not swing with American elections.
The sectoral layer: durable, but without a single mandate
The default regime maps less cleanly onto hiring algorithms, insurance underwriting, public-benefits screening, policing, and healthcare. Here state law can look more like a durable regime than a swinging pendulum. Illinois’s employment rules, for example, have been amended and expanded over more than a decade without federal displacement. Other areas have their own statutory and regulatory frameworks, although their scope and maturity vary considerably by sector and state.
Public distrust of institutions appears to be one driver. Algorithmic tools in these domains increasingly meet demands for bias audits, disclosures, and human-review rights as conditions of use. State laws and existing civil-rights frameworks can impose transparency and anti-discrimination obligations, though the specific requirements are not uniform. Deployment is no longer purely a technical question; it is also a legitimacy question, addressed by compliance departments, regulators, state attorneys general, and courts as much as by engineers.
This may be the most durable layer, but it is not a unified one. State legislatures have democratic mandates, and state attorneys general operate within established legal responsibilities. The problem is that these separate mandates do not add up to a coordinated national framework. Rules emerge from multiple jurisdictions, sometimes addressing similar risks in different ways, while no single institution is responsible for the cumulative architecture.
The courts above all three
After Loper Bright Enterprises v. Raimondo (2024) overturned Chevron deference, federal courts must exercise independent judgment when interpreting ambiguous statutes rather than defer to an agency’s reasonable interpretation simply because the statute is ambiguous. Agency expertise and reasoning can still be persuasive, but they no longer receive Chevron’s mandatory deference. Combined with the major questions doctrine and the open question of whether AI rulemaking has clear congressional authorization in particular areas, this creates additional obstacles for agencies seeking to establish broad regulatory authority. AI rules may be difficult to issue and harder to sustain when their statutory basis is uncertain.
These constraints can cut in more than one direction. They can limit an agency that wants to regulate, and they can also limit an executive that wants to use federal authority to displace state rules. The Colorado stay shows a court halting enforcement of a state law before it took effect; the California ruling shows a court declining to grant the requested preliminary injunction. Neither outcome, by itself, settles the broader legal questions. Courts do not possess a general veto over every regulatory action. They decide specific disputes brought before them, within the limits of jurisdiction, procedure, and the claims presented. But judicial review is an enduring constraint on every layer of the system: it determines which rules can be enforced, how far they reach, and whether the institutions that created them acted within their authority.
What this argument does not claim
Preemption could produce coherence. A single national standard could be simpler for firms than a fifty-state patchwork. But uniformity and effective governance are not the same thing. A national standard could be coherent and still be weak, incomplete, or poorly coordinated with existing sector-specific laws. Durable coherence requires a political coalition willing to establish meaningful obligations and sustain them. The current preemption project is an executive-branch litigation campaign, not a legislative bargain. Preemption without consensus is another swing of the pendulum, one that a future administration can reverse.
A divided-government bargain is the most plausible way out. Trump’s own stated theory is that vetoes lead to deals. Industry wants a national standard, and many Democrats want real federal obligations, so a trade of preemption for baseline duties is conceivable. If that happens, the default regime could begin to give way to a more coherent one. The argument here is that such a bargain is the exception, that it would need to survive more than one administration, and that it would still face judicial review. The Senate voted 99–1 to strike the House reconciliation provision that would have imposed a 10-year moratorium on state AI laws. A future preemption agreement would have a better chance of lasting if it arrived packaged with substantive national standards.
States can be laboratories of democracy rather than sources of compliance cost. Sometimes they are. State experimentation can reveal problems, test solutions, and create models for wider adoption. But when firms build to the strictest state’s standard and apply it nationally, the laboratory can become a de facto national rulemaking process without a corresponding national debate or mandate. The country adopts the laboratory’s output by default, not necessarily because Congress or the public has chosen it.
The independent surge may not realign politics. Most independents still lean toward a party and vote accordingly, and part of the surge reflects reaction to whoever holds power. The claim is not that a new centrist majority is emerging, or that independent identification itself causes congressional instability. It is that looser party attachments, alongside polarization and other structural pressures, are part of a political environment in which durable federal statutes are harder to build.
EU enforcement may be slower and more limited than its formal powers suggest. It probably will be. But it still sets the baseline for EU market access. For firms that choose a common global product architecture rather than regional versions, those requirements can influence design well beyond Europe. That influence is commercially contingent, not automatic, but it does not depend on the outcome of American elections.
The regime without a regime
The United States is not choosing between regulating AI and not regulating it. It is choosing, mostly by failing to choose, who writes the rules by default. Right now that means executive orders that can be reversed, state legislatures that can shape national product requirements by accident, a federal litigation effort that challenges some of those same requirements, courts that determine the legal boundaries, compliance departments that translate public distrust into audits and disclosures, and European regulators whose rules influence any firm seeking access to the EU market.
The costly thing is not political volatility as such. Democracies realign, and realignment can be healthy. The costly thing is the difficulty of forming durable coalitions, the kind that can write a coherent rulebook and be held accountable for it. Until that capacity returns, American AI development may continue largely on its own terms at the federal level, constrained at the edges by state requirements and by European rules that apply to firms selling there. American AI deployment will increasingly be governed by a regime no single institution designed, and for whose cumulative consequences no single institution can be held accountable.
There is a geopolitical cost, too. The United States could remain the world’s leading AI developer while ceding the rule-making layer to institutions that do not share its strategic priorities. When American firms design around European requirements, those requirements reflect European values—precaution over innovation, privacy over security, process over capability. When individual states shape national product requirements by accident, there is no mechanism to ensure those rules align with national security interests or economic statecraft. When successive administrations undo one another’s priorities, foreign regulators gain influence over the practical rules of AI deployment faster than American institutions can reach agreement at home. That outcome is not inevitable, but it is a strategic risk built into the default regime.
The deeper asymmetry is this: American companies may build the most powerful AI systems in the world, but foreign regulators and state attorneys general will decide how those systems can be used. The rules that shape what American AI can do in the world will therefore be written, in significant part, somewhere other than Washington. And the United States may discover that leading in the development of a technology does not guarantee leadership in governing its consequences.