Two House lawmakers are negotiating a bipartisan federal AI bill that would do something the tech industry has wanted for years: preempt state AI safety laws, specifically those targeting model developers in California and New York.

Reps. Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.) are leading the talks, according to Politico, which reported that the emerging bill would also create a federal framework for vetting AI models before release. Think of it as an FDA for AI: pre-market review before the most powerful models can ship.

What Gets Killed

The bill would specifically preempt state AI safety laws that regulate model developers. California's recently passed AI safety law and New York's equivalent would be the primary casualties. Both states passed legislation requiring AI labs to test frontier models for dangerous capabilities before deployment, to report safety incidents, and to maintain certain safety infrastructure.

State laws that do not regulate model developers would reportedly not be affected. That means employment discrimination laws, consumer protection rules, and deployer-focused regulations could survive. The preemption targets the labs, not the users.

The Sunset Trick

Here is the clever part. The negotiators are discussing a sunset provision that would allow states to pass their own AI safety laws again after the federal framework expires. In theory, this means the preemption is temporary. In practice, sunset provisions in federal law almost always get renewed. Once the industry builds compliance infrastructure around the federal standard, the cost of switching back to 50 different state regimes becomes the argument against ever letting the sunset happen.

The Context: 1,200 State Bills and Counting

This comes at a moment of maximum regulatory chaos. Fortune reported this week that there are now over 1,200 AI bills pending across U.S. state legislatures. The 2026 NDAA already banned DeepSeek and High Flyer from defense networks. Colorado just signed a softer replacement of its own landmark AI law after xAI and the DOJ challenged the original. Connecticut's AI bill is moving through committee. The patchwork is becoming ungovernable.

The industry argument is straightforward: AI labs cannot build compliance systems for 50 different state laws. A single federal standard, even an imperfect one, is better than regulatory fragmentation that slows deployment and favors well-funded incumbents who can afford 50 legal teams.

The Counter-Argument

Critics will point out that federal preemption trades real protections for the promise of future ones. California's law exists. New York's law exists. The federal bill does not exist yet. Killing the laws that work to create a framework that might work is a bet, and the tech industry has a track record of lobbying for weaker federal standards precisely to preempt stronger state ones.

The bipartisan nature of the talks makes this more dangerous for AI safety advocates than a purely partisan effort. It is hard to argue against a bill that both parties support. And the "FDA for AI" framing gives it a veneer of rigor that pure deregulation never could.

Who Wins

OpenAI, Anthropic, Google, and Meta win. They get one set of rules instead of 50. The model-vetting requirement is a cost they can absorb. Startups and open-source developers could get squeezed if the vetting process is expensive or slow. And the states that moved first on AI safety lose the regulatory authority they fought to establish.

The talks are ongoing. No bill text has been released. But the fact that bipartisan negotiations are actively happening, with specific state laws named as targets, means the window for state-level AI regulation may be closing faster than anyone expected.

Sources: Politico, Fortune, Transparency Coalition, NPR