The preemption fight: Washington vs the states

Express, conflict, and field preemption

Under the Supremacy Clause, federal statutes displace state law three ways: express (Congress says so in text), conflict (compliance with both is impossible, or state law obstructs federal objectives), and field (federal regulation so pervasive it occupies the field). The AI problem: with no federal AI statute, there is no express text, a voluntary framework creates no conflict — you can comply with the NIST RMF and Colorado — and an empty field cannot be occupied. Deregulatory policy is not preemptive law; the Supreme Court has required a genuine federal enactment.

The Spending Clause and the coercion limit

Congress (and agencies exercising delegated spending discretion) may attach conditions to federal funds — the EO 14365 lever. But conditions must be unambiguous, related to the program’s purpose, and non-coercive. In NFIB v. Sebelius (2012) the Supreme Court struck down threatening states’ entire existing Medicaid funding as ‘a gun to the head’. Expect exactly this fight over conditioning broadband money on AI-law non-enforcement: is that condition related to broadband, and is the amount at stake coercive? Also contested: whether an agency may invent such conditions without Congress writing them into the appropriation.

Dormant Commerce Clause

Even with Congress silent, states may not discriminate against or unduly burden interstate commerce. The argument against state AI laws: a model trained once and served nationwide cannot vary by state, so one state effectively regulates the nation (the ‘extraterritoriality’ theory). The counter: National Pork Producers v. Ross (2023) upheld California’s Proposition 12 against precisely this attack — states may regulate in-state effects even when compliance costs propagate nationally. This is the strongest doctrinal battleground for the DOJ task force, and its outcome is genuinely uncertain.

First Amendment

Two live theories: compelled speech — mandatory disclosures (‘this content is AI-generated’, published model policies) compel companies to speak, tested under Zauderer’s lenient standard for factual commercial disclosure or stricter scrutiny if courts see them as controversial; and model outputs as protected speech — if AI outputs are speech, laws restricting them face scrutiny. Courts have only begun sorting these claims; early rulings point in different directions.

Can the feds actually knock out state AI law X?

Interactive decision tree — outcomes:

• State law falls

  With a real federal statute plus express or conflict preemption, the Supremacy Clause does its work. This is the scenario EO 14365’s legislative framework is trying to create — and the one that does not exist as of September 2026.

• Both laws stand

  Federal floors with state add-ons are the American default (think privacy, employment law). Companies comply with both — plan for the strictest applicable rule.

• Pressure works — politically, not legally

  Lawful funding conditions don’t invalidate the state law; they make enforcing it expensive. States choose between money and policy. Watch for exactly this dynamic reshaping state drafting: carve-outs and disclosure-style duties instead of hard mandates.

• Condition likely unconstitutional

  Unrelated or coercive conditions fail NFIB v. Sebelius — and conditions never authorized by Congress face separation-of-powers attack on top. Expect states to sue rather than fold.

• Serious constitutional vulnerability

  Discrimination against interstate commerce or direct regulation of wholly out-of-state conduct is where dormant Commerce Clause challenges win. State AI laws drafted with in-state nexus requirements try to stay out of this box.

• Challenge faces an uphill climb

  National Pork Producers v. Ross (2023) blessed state laws whose compliance costs ripple nationwide, absent discrimination. Not a guarantee — the doctrine is contested — but the state starts ahead.

• Unsettled — watch the courts

  Factual commercial disclosures usually survive under Zauderer; restrictions on model outputs raise harder questions courts have only begun answering. No stable rule exists as of September 2026 — check current litigation before advising.