U.S. wavers on AI regulation

AI policy in the U.S. just got more confusing — and more concrete. On May 5, the Commerce Department’s AI testing arm signed new agreements with Google DeepMind, Microsoft, and xAI so the government can evaluate frontier models before the public sees them. But the broader White House line is still pretty light-touch: build a national framework, stop the states from making a patchwork, and avoid heavy new federal restrictions. ### What happened this week? The immediate news is the NIST announcement. CAISI — short for the Center for AI Standards and Innovation — said the three companies will let it run pre-deployment evaluations, post-deployment assessments, and related research on advanced models. The point is national security testing, not consumer protection in the broad sense. CAISI said it has already completed more than 40 evaluations, including on unreleased systems. (nist.gov) ### What is CAISI actually doing? Basically, CAISI is becoming the government’s technical checkpoint for frontier AI. The agreements let developers share models, sometimes with safeguards reduced or removed, so evaluators can probe dangerous capabilities more directly. Testing can happen in classified environments, and other agencies can join through an interagency task force. That is more hands-on than a vague voluntary pledge — but it is still cooperative, not a licensing regime. (nist.gov) ### So is Washington tightening AI rules? Yes and no — that’s the wobble. The administration is clearly building more federal capacity to inspect powerful models. But its March 20 legislative framework does not read like a push for an EU-style AI rulebook. It leans toward targeted rules on children, scams, infrastructure, IP, and national security, while stressing innovation, lighter permitting, and a single national approach instead of lots of state laws. (nist.gov) ### Why does the White House keep talking about one national framework? Because the administration sees state-by-state AI law as the bigger threat to industry. The March framework explicitly argues that fragmented state regulation could hinder U.S. competitiveness, and later White House materials frame federal policy as a way to preempt that patchwork. In plain English — Washington wants enough federal policy to keep control in Washington, but not so much that model developers feel boxed in. (whitehouse.gov) ### Why are companies okay with model testing? Because this version is collaborative and narrow. The companies get a government partner focused on national security risks, measurement, and best practices — not a standing regulator with power to block launches across the board. That can help labs show they are being responsible without signing up for a much tougher system later. That last part is an inference, but it fits the structure of the agreements and the administration’s broader posture. (whitehouse.gov) ### What does Europe have to do with this? A lot. While the U.S. is building a cooperative testing system, Europe is still fighting over how hard and how fast to enforce the AI Act. Current omnibus debates could delay key obligations for high-risk systems to December 2027 or even August 2028, while also creating loopholes for systems already on the market. So the odd twist is that both sides of the Atlantic are wavering — just in different ways. (nist.gov) ### What is the real takeaway? The U.S. is not doing “anything goes.” It is building federal AI oversight capacity right now. But it is also not choosing broad, hard-edged regulation. The model taking shape is softer and more strategic — test frontier systems, centralize authority, and keep the brakes light unless a specific harm or security risk forces a harder turn. (nist.gov) (techpolicy.press)