White House weighs model reviews

AI policy in Washington just got more concrete. The Trump White House is now discussing whether some new AI models should face government review before the public gets them. That is a much(usnews.com)gap was obvious — lots of talk about leadership, not much clarity on what happens when a frontier model is powerful enough to create real security or safety problems. On May 4, that started to change. (usnews.com) ### W(whitehouse.gov)ls. The discussions also include an executive order that could create an AI working group bringing together government officials and tech executives, basically as a mechanism for testing, consultation, and escalation before a model goes wide. It is still a deliberation, not a final rule. But the fact that the administration is even entertaining model reviews is the news. (nytimes.com) ### Why is that a big shift? Because this administration spent 2025 moving in the other di(usnews.com)peed, infrastructure, exports, and cutting regulation that might slow private companies down. A pre-release review system would not erase that strategy, but it would add a clear gate in front of public deployment. (whitehouse.gov) ### Why review the model before release? Because once a powerful model is public, the hard part is no longer invention — it is containment. Frontier systems can be p(nytimes.com)ers evade safeguards. Pre-release review is the policy version of inspecting an aircraft before takeoff rather than after the crash. The government seems to be circling that logic now. (nytimes.com) ### Would this look like licensing? Not necessarily. What has surfaced so far sounds more like a review-and-testing layer than a blanket licensing regime for all AI. The practical quest(whitehouse.gov)tied to certain risk thresholds. Those details are still missing, and they are everything. (nytimes.com) ### Why mention a working group? Because the White House may want a faster, more flexible structure than formal rulemaking. A working group can pull in companies building the biggest models, national security officials, and technical agencies without first wr(nytimes.com)ormation, and pressure firms into common release practices — even before Congress acts. That last part is an inference, but it fits the structure being discussed. (bloomberg.com) ### Who would feel this first? The obvious targets are the companies training the most c(nytimes.com)ure would land upstream — on model developers that can afford giant training runs and are already doing internal safety testing. For enterprise buyers, that could eventually make audits, evals, logging, and release documentation feel less optional and more like table stakes. (nytimes.com) ### What is still unresolved? Almost all of the hard parts. There is no public final standard yet for what triggers review, what happens if a mode(bloomberg.com)s the whole story — move too slowly and the White House fears losing the race, move too fast and it owns the fallout from a model that should have been stopped. (usnews.com) ### Bottom line This is still a proposal, but it is not a trivial one. If the White House follows through, the federal government would b(nytimes.com)r the most powerful models. (usnews.com)