Tech giants push federal AI preemption as states split

A growing bloc of Big Tech firms is steering Washington toward preempting state AI rules even as states pass their own—an uneasy tug-of-war that sharpened this summer, according to a Bloomberg report on industry efforts to block state AI rules. The claim: a national standard will avoid a 50-state compliance maze. The reality: states aren’t waiting.

The coordination is visible. Republicans failed in June to attach a 10-year moratorium on state AI enforcement to a tax bill, but the White House later signaled that federal agencies should avoid funding states with “unduly restrictive” AI regulation. Meanwhile, more than 1,000 state AI bills landed this year, though only several dozen would truly constrain developers or deployers. Most are gestures. Some bite.

The push is also strategic, not blanket opposition. Tech companies say they can live with rules that govern use—how customers deploy AI—over rules that govern development, which shape how models are built and trained. That distinction shifts compliance downstream. It matters.

And the clock is ticking. Colorado’s first-in-the-nation anti-discrimination law, heavily contested, is being reworked in a special session and delayed, but it remains the closest thing to a comprehensive U.S. AI regime. Pressure works.

Washington’s new leverage point

By tying federal dollars to “friendly” jurisdictions, the administration gave industry a new wedge against the patchwork problem. It’s soft power but real. That leverage matters.

Trade groups and investors frame the ask as common sense: align on national baselines, target harmful uses, and spare frontier R&D from conflicting state mandates. States hear something else: a preemption land-grab while Congress stalls. Both readings can be true.

This is where language choices do heavy lifting. “Regulate use, not development” sounds small. It redraws the line of accountability.

Two state templates are emerging

The first template bans or sharply restricts AI in mental health settings. Nevada’s AB 406 and a similar Illinois measure bar developers from claiming their systems can provide professional therapy or simulate conversational therapy, and they block schools from outsourcing counseling to AI. The statutes carve out “self-help” materials, are keyed to “professional” services, and talk about systems “specifically programmed” for therapy.

Those phrases create escape ramps. General-purpose models can argue they’re not “specifically programmed.” Self-help disclaimers abound. The result is odd: unless companies aggressively monitor user prompts and outputs, compliance becomes guesswork. It won’t stop use.

The second template targets frontier transparency. California’s SB 53 and New York’s RAISE Act focus on the largest developers with safety-and-security protocols, critical-incident reporting, and (in California’s case, from 2030) third-party audits. They regulate entities, not model features, leaving technical choices flexible while compelling disclosures. That’s pseudoregulation: public oversight of private guardrails.

Both approaches avoid grand AI “comprehensives.” They’re narrower. They still set precedent.

Colorado’s cautionary tale

Colorado’s 2024 law against “algorithmic discrimination” catalyzed this year’s reckoning. The governor called for revisions almost immediately. The sponsor agreed tweaks were necessary. In a special session this week, lawmakers advanced one bill that shifts obligations toward developers and delays effect by months, while a competing bipartisan bill attempts a longer delay to 2026 with few substantive changes. The committee maneuvering has been unusually aggressive.

Why it matters: first-mover disadvantage. When a state goes first on a fast-moving technology, it inherits all the implementation pain and most of the lobbying heat. Later states calibrate around the backlash. Colorado now sits in that vice.

The economy, unsurprisingly, is the cudgel. Business groups argue the existing statute would be onerous and expensive; consumer advocates counter that transparency and redress are the point. Neither is entirely wrong. The politics are the policy.

Enforcement meets reality

Mental-health bans protect licensed professionals and public school staff from technological encroachment in the name of safety. They also nudge companies toward heavier content surveillance to avoid “implicitly” offering therapy, an awkward fit for general-purpose systems trained on vast public text that often includes therapeutic language. Geography filters, disclaimers, and stricter refusal regimes will rise. Users will route around them.

Frontier transparency, for its part, builds path dependency around a specific risk set: catastrophic cyber, CBRN misuse, model deception, and loss of control. Those are legitimate. They are also infrequent compared with today’s commonplace harms: fraud, spam, disinformation, and unfair denials in credit or hiring. Audits tend to drift toward checkbox compliance over time. So might this.

One more blind spot: robotics. If “frontier developer” soon means AI embedded in machines, do SB 53-style disclosures fit factories and homes? Laws written for digital agents will creak when hardware enters the chat. Future sessions will patch. The patches will differ.

The bigger market logic

A patchwork favors giants. Large firms can amortize compliance across products and states and wield relocation threats credibly. Startups cannot. Federal preemption would neutralize that leverage and simplify playbooks. That’s the sell in Washington. It’s persuasive inside the Beltway.

But states legislate because Congress won’t. And every cycle without a federal bill gives statehouses room to experiment, retreat, and try again. The center of gravity is still moving.

Why this matters

• Patchwork vs. preemption is now the core fight. Early state movers absorb the costs and political heat while industry leverages federal funding guidance to slow or reshape them.
• Today’s bans set tomorrow’s defaults. Mental-health and frontier-safety rules will harden product choices—monitoring, refusals, audits—even if users and risks evolve faster than statutes.