The Only Binding AI Rules Are the States' — and Both Sides Are Trying to End Them
The only rules that actually bind what AI models do are the states' — and by November, the labs and Washington were each trying to end them their own way.
Right now, the only rules that actually bind what the most powerful AI models do are written in state capitals. In New York, the RAISE Act — which required the companies building frontier models to register with the state by November — keeps a seventy-two-hour clock on reporting any critical safety incident, and demands quarterly assessments of catastrophic risk [1]. In California, an executive order is putting a kill switch on the drawing board, directing an expert panel to design a way to shut a frontier system down after a breach and research showing models resist shutdown commands as much as 97 percent of the time [1]. Across the country, 85 such laws passed in 27 states this year [2][1]. Companies build to the strictest of them, so these state-shaped rules already function as national ones in practice [3] — and they collide, New York's seventy-two-hour clock not lining up with California's own reporting rules [1]. Nobody designed this patchwork. It accumulated. The companies want it gone, and one of them said so in a lawsuit.
Government regulation that is applied at the state level in a patchwork across the country can have the effect to hamper innovation and deter competition in an open market. — xAI
xAI filed that suit in April to void Colorado's algorithmic-discrimination law for high-risk systems [4]. But the more revealing arc is the one that ran from August to October. In August, Anthropic's Dario Amodei asked the federal government to test frontier models before public release [5]. In September, OpenAI asked in writing for mandatory reporting of safety incidents [6]. In October, Anthropic drafted a federal safety bill whose binding provisions reach all the way to Sacramento: Commerce could ask a Washington judge to block any model posing a foreseeable, imminent risk of more than a billion dollars in damages or fifty deaths — and the bill would override certain state laws, including California's, while barring other federal agencies from halting development [7]. The administration kept its own conduct rules voluntary — the September 29 accord's "independent oversight" is still undefined [8] — and pointed its executive hand outward instead. Executive Order 14365 challenges state AI restrictions that conflict with federal policy [3], and the Justice Department is expected to challenge California's kill-switch order [1]. OpenAI, for its part, has endorsed that same California order as a step toward national safeguards.
an important step toward adaptable national safeguards — OpenAI
Congress has the bills but not the settlement: the FRONTIER Act and the AI Risk Evaluation Act sit stalled [9], and last week two dozen House members answered the year's demands with a new caucus, another forum rather than a rule [10]. But the states' binding power isn't only the model rules. It is also a lever no safety bill touches: over buildings. Newly elected Democratic governors in Ohio, Georgia, Texas, and New Mexico are moving to pause data-center construction [11]. Oklahoma City's council is weighing an extension of its construction moratorium, tied to water limits and community-benefits terms [12]. Siting is a different lever from the model rules that both the labs' bill [7] and the expected Justice Department suit [1] are aimed at. The governors' letter out this week tries to settle the map with a trade [13]. Washington takes frontier safety — the independent federal reviews of powerful models and mandatory reporting of serious incidents that the governors say the federal government is best placed to run [13]. The states keep what their own binding power actually does: consumer protection, discrimination cases, children, workforce [13]. If the trade is taken, the clock and the switch change addresses. Until it is, the binding rules keep their current ones — and both of this year's attempts to move them, the lab-drafted bill and the expected suit, are aimed at those same addresses. The governors have made the offer. No one has accepted it.
- 1. California and New York Launch State-Level AI Safety Mandates
- 2. Trump Appoints AI Czar as US Treats AI as Public Safety Risk
- 3. US State AI Laws Create Fragmented Compliance Landscape
- 4. xAI Sues Colorado to Block AI Discrimination Law
- 5. AI Firms Call for Federal Oversight of Frontier Models
- 6. OpenAI Urges Congress to Mandate National AI Safety Rules
- 7. Anthropic Pitches AI Safety Bill to U.S. Lawmakers
- 8. White House AI Safety Accord Lacks Independent Oversight Standards
- 9. U.S. Pursues Federal AI Framework Amid State Resistance
- 10. House Lawmakers Form Bipartisan AI Safety and Innovation Caucus
- 11. Democratic Governors Propose Data Center Pauses After Midterms
- 12. Oklahoma City Council Votes on Data Center Moratorium Extension
- 13. U.S. Governors Urge Federal AI Safety Regulations