The U.S. is studying whether to require emergency mechanisms to stop advanced AI models, as experts warn that a simple off switch may not be enough to halt a system that becomes too powerful.