New research suggests Chinese state media and authoritarian speech restrictions can shape the answers produced by leading U.S. AI models