AI rules and pauses

OpenAI got hit from three directions in two days: a state prosecutor in Florida, lawmakers in Illinois, and the power grid in Britain. The company is being asked what its chatbot does to children, what harms it should pay for, and whether it can even afford the electricity needed to keep expanding. (cbsnews.com) (wired.com) (money.usnews.com) On April 9, Florida attorney general James Uthmeier said his office had opened an investigation into OpenAI over alleged risks to minors, public safety, privacy, and national security. He also pointed to a possible link between ChatGPT and a 2025 shooting at Florida State University, which pushed the probe out of the usual “online safety” lane and into criminal-law politics. (cbsnews.com) (techcrunch.com) That kind of state investigation matters because there is still no single federal rulebook for advanced artificial intelligence in the United States. So attorneys general, state legislatures, and courts are each testing different ways to police the same technology, one case at a time. (cbsnews.com) (legiscan.com) At almost the same moment, Wired reported that OpenAI testified for an Illinois bill that would sharply limit when model makers can be sued for “critical harm.” The bill covers extreme cases like death or serious injury of 100 or more people, or at least $1 billion in property damage, and would make it harder to hold the lab itself liable if someone used its model to cause the damage. (wired.com) That is the legal fight underneath a lot of artificial intelligence policy right now: is a model maker closer to a car company, which can be blamed for a defective product, or closer to a phone company, which usually is not blamed for what a customer says on the line. OpenAI is arguing for the second bucket in the worst-case scenarios. (wired.com) California moved in the opposite direction last year with Senate Bill 53, the Transparency in Frontier Artificial Intelligence Act, which requires the biggest developers to publish safety frameworks and spell out how they manage catastrophic-risk scenarios. That law does not settle liability, but it shows states are already building different pieces of an artificial intelligence rulebook instead of waiting for Washington. (legiscan.com) (ai-analytics.wharton.upenn.edu) Then there is the less glamorous bottleneck: power. Reuters reported on April 9 that OpenAI paused its main British data-centre project because of high energy costs and an unfavorable regulatory environment, stalling a multibillion-pound plan that Britain had hoped would strengthen its domestic artificial intelligence capacity. (money.usnews.com) (cnbc.com) A data centre is just a warehouse full of computers, but for artificial intelligence those warehouses behave more like steel mills than office buildings. If the local price of electricity is too high, or permits take too long, the model does not care how friendly the press release was. (money.usnews.com) (datacentrereview.com) Britain had pitched itself as a place to host “sovereign” computing power, meaning compute capacity located inside the country rather than rented abroad. OpenAI’s pause is a reminder that artificial intelligence policy is not only about safety memos and lawsuits; it is also about substations, land, and utility bills. (cnbc.com) (datacentremagazine.com) Put those three moves together and the picture gets clearer. OpenAI is facing a patchwork world where one state prosecutor can open a child-safety probe, another state can debate shielding labs from mass-harm lawsuits, and another country can slow expansion simply by making power too expensive. (cbsnews.com) (wired.com) (money.usnews.com) That is what artificial intelligence regulation looks like in 2026: not one grand law, but a stack of local fights over children, liability, planning, and electricity. The companies building the models are learning that the future of artificial intelligence will be written as much in attorneys general letters and power contracts as in code. (cbsnews.com) (wired.com) (money.usnews.com)