The past week produced a strange kind of governance weather: three separate attempts to regulate artificial intelligence, at three different scales, all moving simultaneously and none of them in step with the others. Colorado passed a law. The White House had an internal knife fight. And OpenAI pitched something that sounds like a United Nations for AI, timed to a presidential trip to Beijing.
Start at the state level, where coherence has, improbably, arrived. Colorado's SB 189 passed Tuesday with votes of 57-6 in the House and 34-1 in the Senate, ending two years of deadlock over what to do with the state's 2024 Colorado AI Act, the first law of its kind in the US. That earlier law had required companies to prevent discrimination when AI was used in consequential decisions about loans, healthcare, housing, or insurance. It was signed with the promise it would be revised before taking effect. But revision proved harder than anyone expected: a special legislative session stalled out, and last month the Department of Justice filed a legal challenge that temporarily blocked enforcement.
What finally broke the logjam was a six-month working group appointed by Governor Polis, which produced 11 drafts before landing on a proposal. SB 189 takes a lighter touch than its predecessor, narrowing the scope of what counts as "consequential" AI and softening requirements on developers, but digital rights advocates say key safeguards survive. "This bill is probably going to be used as a model in other states," said Loren Furman, CEO of the Colorado Chamber of Commerce and a working group member. Given how few states have passed substantive AI consumer protection laws, she may be right.
Meanwhile, the federal picture looks nothing like Colorado's bipartisan tidiness. A reporting roundup from security newsletter Risky Business News captured what the White House's own sources called a "knife fight" between the Commerce Department and national security aides over who gets to vet AI models before release. Commerce already has a head start: its Center for AI Standards and Innovation (CAISI, formerly the US AI Safety Institute) had built testing infrastructure and secured voluntary pre-release agreements with Google, Microsoft, and xAI. On Friday, CAISI quietly took its own announcement page down, reportedly due to "sensitivity" from the White House. The National Cyber Director has separately proposed a new evaluation center within the Office of the Director of National Intelligence. No one yet has jurisdiction.
The intelligence community angle makes some sense: the NSA and its peers have genuine expertise in national security implications of AI models. The problem, as Risky Business noted, is that AI's potential harms span biology, chemistry, economics, and labor -- domains that sit well outside the intelligence community's core competency. There is no single agency that contains all the needed expertise, and the current scramble reflects that reality. Lobbyists describe the internal situation as having "no clarity" because "different factions within the White House have different views."
Into this vacuum stepped OpenAI, with a proposal that skipped the domestic argument entirely. Speaking as President Trump arrived in Beijing, OpenAI's Vice President of Global Affairs Chris Lehane said the company would support a US-led global AI governance body that includes China. The model: the International Atomic Energy Agency, which manages nuclear safety standards across member states. The mechanism: connecting the Commerce Department's CAISI with AI safety institutes being built elsewhere in the world. Whether the Trump administration would support any framework that gives China a seat at the governance table remains unclear; White House officials have previously indicated they oppose worldwide governance over AI.
What stands out, reading these three stories together, is the inversion of the usual political logic. States are supposed to move slowly; Colorado moved faster than Washington. Companies are supposed to resist regulation; OpenAI just proposed more of it. The federal government is supposed to set frameworks; it is busy arguing about who owns the table. AI regulation may end up being shaped less by any single authority than by whichever level of government manages to act first.