OpenAI just pitched letting states write the rules first, then scaling the winners nationwide — a playbook that worked for securities law and could determine whether AI governance happens at all.
The Summary
- OpenAI proposes "reverse federalism" for AI regulation: states experiment with safety frameworks, federal government harmonizes what works
- California, Colorado, and Texas already testing different models for AI testing, disclosure, and liability
- The real game: whoever writes the baseline state laws writes the national standard by default
The Signal
OpenAI is doing something clever here. Instead of waiting for Congress to figure out what AI safety means, they're backing a state-by-state approach where California tests disclosure rules, Colorado pilots pre-deployment testing requirements, and Texas experiments with liability frameworks. The best ideas get adopted by other states. The federal government eventually harmonizes the patchwork. This is exactly how securities regulation evolved after the 1929 crash, when state blue sky laws became the template for the SEC.
The timing matters. Federal AI legislation has stalled for two years. The EU AI Act is already in force. China's algorithm registry is live. The US needs a governance model that can move faster than congressional committees but slower than a single executive order that gets reversed every four years.
"State experimentation gives us real-world data on what safety measures actually work, not just what sounds good in a hearing room."
Here's what's actually being tested right now:
- California: Mandatory red-teaming for frontier models before public release
- Colorado: AI impact assessments for high-risk applications in hiring, lending, healthcare
- Texas: Strict liability for autonomous systems that cause physical harm, zero liability for conversational AI
The differences matter. If California's framework becomes the default, every lab building frontier models will need state approval before launch. If Texas wins, the rules stay light unless your AI drives a car into someone. OpenAI isn't neutral here. They're explicitly backing the California model while calling it "federalism."
The Implication
Watch which state frameworks get adopted by multiple states in the next 12 months. That's your preview of federal policy. If you're building AI products, you're already designing for the strictest state standard, which means California's requirements become everyone's requirements.
The real question: does this actually create safety standards, or does it just let the biggest labs write the rules through state capture before anyone else can compete. OpenAI calling this democratic doesn't make it so.