White House studies AI vetting
- Kevin Hassett said on May 6 the White House is studying an executive order that could force frontier AI models through government safety review. - The clearest clue is NIST’s CAISI office — on May 5 it expanded pre-release testing deals with Google DeepMind, Microsoft, and xAI. - That would mark a shift from last year’s pro-speed AI plan toward a gatekeeping role focused on security.
The White House is talking about something much more muscular than the usual AI principles-and-frameworks stuff. It is studying an executive order that could make the federal government review powerful AI models before they are released, with Kevin Hassett comparing the idea to how the FDA handles drugs. The reason is simple — frontier models are no longer just chatbots. They can expose cyber weaknesses, automate misuse, and get deployed into business and government systems before anyone outside the company has really stress-tested them. (bloomberg.com) ### What actually changed this week? The immediate news came in two parts. On May 5, NIST’s Center for AI Standards and Innovation — CAISI — said it had expanded agreements with Google DeepMind, Microsoft, and xAI to run pre-deployment evaluations and related research on frontie(bloomberg.com)etting future systems before public release. (nist.gov) ### What is the White House trying to build? Basically, a gate before launch. The idea under discussion is not just voluntary red-teaming or companies publishing their own safety cards. It is a formal review process that could examine whether a model creates national security(nist.gov)dvanced systems. (politico.com) ### Why compare AI to an FDA drug? Because the administration wants a metaphor people instantly understand — you do not ship a powerful thing into the world first and ask safety questions later. Hassett’s point was that advanced models may create vulnerabilities of their own, so the gov(politico.com)I and medicine are technically similar. It means the White House is signaling a pre-clearance mindset. (bloomberg.com) ### Why is CAISI at the center of this? CAISI already has the machinery. NIST said the office has completed more than 40 evaluations, including on unreleased state-of-the-art models, and that developers often provide versions with safeguards reduced or removed so the government (bloomberg.com)vernment for testing, collaborative research, and best-practice development around commercial AI systems. (nist.gov) ### So is this mandatory now? No — and that is the important catch. Right now, the expanded CAISI arrangements are agreements with companies, not a universal legal licensing regime. The executive order is still being studied, and the White House has not announced final policy text. But the direction of travel is pretty clear: voluntary testing is becoming the staging ground for something more formal. (politico.com) ### Why is this a bigger shift than it sounds? Because the administration’s 2025 AI Action Plan leaned hard toward speed, infrastructure, exports, and removing barriers to deployment. It talked about winning the AI race and cutting rules seen as slowing development. A pre-release vettin(politico.com)whitehouse.gov) ### What would companies worry about? Delay, discretion, and visibility. If the government can test models before launch, labs may have to reveal more about capabilities, failure modes, and internal safeguards earlier than they want. And if the review standard is fuzzy, companies will worry that “(whitehouse.gov) from the fact that the White House is considering a process that could potentially block release. (politico.com) ### Bottom line? This is the first real sign that Washington may want to move from advising frontier AI labs to screening them. If that happens, the U.S. would still be trying to win the AI race — but with a federal checkpoint at the starting line. (whitehouse.gov)
