What changed in the rules for AI models
The U.S. administration has officially signaled to developers: open models with public weights will no longer undergo voluntary safety testing. According to Reuters, a corresponding notice has already been sent to companies working in the field of artificial intelligence. The decision means that regulation will now become more targeted, focusing on products that do not disclose their code.
In other words, the government has changed its approach to oversight: instead of vetting all systems indiscriminately, checks will now apply only to “closed” models whose source code remains inaccessible to external auditors. According to The New York Times, this was reported by three independent sources familiar with the situation. It is a voluntary procedure, not legally mandatory, but one the industry had come to regard as the standard.
The logic behind the decision
The choice to review only closed models looks logical from a regulatory standpoint. Open weights can be studied, analyzed, and fine-tuned by the community, so the risks of using such models are partly offset by their transparency. Proprietary systems are another matter—no one but the developer knows exactly how the algorithm is built and what data it processes.
Moreover, voluntary testing has always been a soft-power tool: companies agreed to it to demonstrate good faith. Now the administration appears to have decided to lift that burden from those who openly publish models and to focus resources on “black boxes.” This can be understood as a step toward deregulation, while retaining control over the most opaque developments.
What this means for the industry
For open-model developers, the news primarily means less bureaucracy and faster time to market. They no longer need to prepare for lengthy reviews, agree on safety metrics, or wait for regulator approval. This could accelerate the emergence of new open-weight solutions and intensify competition with closed platforms. At the same time, responsibility for security is effectively shifted to the companies themselves and the community.
On the other hand, closed models will have to go through the procedures on their own, which could slow down their adoption. In the future, we will likely see a market split: open development will advance more freely, while closed systems will remain under close government scrutiny. The question remains how effective voluntary checks without mandatory status can be, but for now this is the scheme chosen in the U.S.
Overall, the decision sets a precedent: AI regulation is becoming more differentiated, taking into account the specifics of model architecture. Whether this will stimulate innovation or create blind spots in safety remains to be seen. But one thing is already clear: the era of uniform requirements for all AI systems is coming to an end.
